<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Eni digiTALKS - Medium]]></title>
        <description><![CDATA[Eni’s channel on AI, data science, design, innovative digital tools, methodologies and much more. We share our experiences, knowledge and insights to start a fruitful discussion with the community. Let’s engage in a lively exchange on the exciting world of digital technologies! - Medium]]></description>
        <link>https://medium.com/eni-digitalks?source=rss----fe373d427779---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Eni digiTALKS - Medium</title>
            <link>https://medium.com/eni-digitalks?source=rss----fe373d427779---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 08:11:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/eni-digitalks" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[From Maintenance Reports to Operational Insights: AI-driven tools for Renewable Energy on Databricks]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/eni-digitalks/from-maintenance-reports-to-operational-insights-ai-driven-tools-for-renewable-energy-on-databricks-052b7e0fc1ae?source=rss----fe373d427779---4"><img src="https://cdn-images-1.medium.com/max/756/1*XxmFPctxhDKUIxiktWklAQ.png" width="756"></a></p><p class="medium-feed-snippet">Maintenance reports from O&amp;M suppliers, companies responsible for the operation and maintenance of solar and wind renewable energy plants&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/eni-digitalks/from-maintenance-reports-to-operational-insights-ai-driven-tools-for-renewable-energy-on-databricks-052b7e0fc1ae?source=rss----fe373d427779---4">Continue reading on Eni digiTALKS »</a></p></div>]]></description>
            <link>https://medium.com/eni-digitalks/from-maintenance-reports-to-operational-insights-ai-driven-tools-for-renewable-energy-on-databricks-052b7e0fc1ae?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/052b7e0fc1ae</guid>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[renewable-energy]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Maria Vallarelli]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 07:26:25 GMT</pubDate>
            <atom:updated>2026-10-20T07:26:24.444Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[From Black Boxes to Equations: The Practical Magic of Equation Discovery with AI]]></title>
            <link>https://medium.com/eni-digitalks/from-black-boxes-to-equations-the-practical-magic-of-equation-discovery-with-ai-d02e6aba13cf?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/d02e6aba13cf</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Roberto S. Pinna]]></dc:creator>
            <pubDate>Sun, 28 Jun 2026 13:55:07 GMT</pubDate>
            <atom:updated>2026-06-28T13:55:06.436Z</atom:updated>
            <content:encoded><![CDATA[<p>Surrogate models buy you speed, inverse problems reveal hidden constants — but you may still lack the governing law. Emerging AI tools can extract an interpretable formula you can publish, verify, and deploy and they do it by evolving math like a species: accuracy vs simplicity on a Pareto front.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xEeLb-mKw2bUfiSO81QbVQ.png" /><figcaption>Image generated by AI</figcaption></figure><h3><strong>The Limit of Surrogate Models</strong></h3><p>As data scientists, we are taught a golden rule: let the data speak for itself. One can feed thousands of rows into a <strong>surrogate AI model</strong>, minimize the loss function, and watch the validation metrics soar. But in the real world of engineering and science, a high accuracy score isn’t enough.</p><p>Imagine you are tasked with predicting the temperature of a chemical reactor. You train a surrogate regressor model on historical data, and within the bounds of that data, it interpolates beautifully. But ask that same model to predict what happens if the temperature spikes 50 degrees higher than ever recorded — a critical safety scenario — and it might confidently predict that the reactor will end up fine.</p><p>Why? Because a standard AI models has no inherent concept of thermodynamics. It doesn’t know about conservation of energy or reaction rates. It effectively memorizes curves. Pure machine learning is a master of interpolation, but when forced to extrapolate into the unknown, it often fails disastrously because it lacks the underlying physical laws that govern the system.</p><p>This is the <em>Black Box</em> problem. We have built models that are incredibly predictive but intellectually empty. To move from simple prediction to true understanding — and to trust our AI with critical physical systems — we need models that don’t just fit data, but actually uncover the rules behind it.</p><p>The fastest path from <strong>a model that predicts</strong> to <strong>a model you can inspect</strong> is to turn the problem around: instead of simply mapping outputs from inputs, we ask what hidden causes produced what we measured. That is the domain of <strong>inverse problems.</strong> When we don’t even know the right mathematical form of the law we’re trying to infer, we need something stronger.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uXWowMZHCOaXItTBR6shSg.png" /><figcaption>Reactor thermal model example. Image generated by AI.</figcaption></figure><h3><strong>The Inverse Problem: Playing Detective</strong></h3><p>Data driven models are incredible tools, but they rely on one massive assumption: <strong>you have a complete picture of the domain.</strong></p><p>In a <em>forward problem</em>, you know the rule and you compute the outcome. You know the mass of a pendulum, the length of its string, and the coefficient of air friction; you plug them into the equations and predict where the pendulum will be in ten seconds.</p><p>In an <strong>inverse problem</strong>, you flip the arrow of reasoning: you observe outcomes (data) and you infer the unknown causes (e.g. parameters, boundary conditions, sources, or latent fields). You still assume the <em>structure</em> of the equation is correct — but some pieces of it are unknown, and the data must <em>solve for them</em>.</p><p>Imagine you are handed a shaky, low-resolution video of a pendulum swinging in a sealed chamber. You can track the coordinates of the pendulum over time, but the data is sparse — maybe you only get one frame every few seconds. Your task is to figure out the exact invisible friction coefficient inside that chamber.</p><p>If you try to solve this with pure machine learning, you will likely fail. For example, a standard neural network can fit a curve to the coordinates of the swinging pendulum, but it cannot extract the physical friction coefficient from that curve. It doesn’t know what friction <em>is</em>.</p><p>Physics-Informed AI turns this paradigm upside down. We can use the differential equations of classical mechanics not just to guide the network, but to actively hunt for missing parameters.</p><h4><strong>The AI Detective Architecture</strong></h4><p>Here is how to tackle an inverse problem using a physics-informed architecture.<br>Instead of just treating the neural network’s internal weights and biases as trainable parameters, <strong>we declare the unknown physical constant (our friction coefficient) as a trainable parameter, too.</strong></p><p>We then set up a dual-objective loss function:</p><p>1. <strong>Data Loss (The Observation):</strong> Does the neural network’s predicted pendulum path match the sparse coordinates we extracted from the video?</p><p>2. <strong>Physics Loss (The Interrogation):</strong> Does the network’s predicted path, <em>combined with its current guess for the friction coefficient</em>, mathematically obey the <strong>classical pendulum equation</strong>?</p><p>As the neural network trains via gradient descent, something magical happens. The backpropagation algorithm isn’t just updating the hidden layers of the network to make the curve look good. It is simultaneously tweaking that unknown friction parameter to reduce the dual-objective loss.</p><p>By the end of training, your model hasn’t just learned to interpolate the missing pendulum positions. It has played detective. It looked at a few sparse dots on a graph, consulted the laws of classical mechanics, and mathematically deduced the forces acting on the system.</p><p>This isn’t limited to pendulums, other examples can be the inference of thermal conductivity for materials or fluid viscosity using field data.</p><p>We have learned how to interrogate physics for unknown physical parameters (Inverse Problems). But what if we want the AI to give us the ultimate prize: a novel, human-readable equation?<br>That requires a completely different kind of machine learning.</p><h3><strong>Symbolic Regression: AI That Writes Equations</strong></h3><p>So far, we played detective, using inverse modeling to find a hidden friction coefficient for a swinging pendulum. But that process still relied on one massive, fundamental assumption: <strong>we already knew the pendulum equation.</strong> Inverse problems are powerful, but they are brittle. If you assume the phenomenon follows a linear law (like y = mx + c) and solve for m and c, but the reality is actually non-linear, your inverse problem will fail silently. You will find the best parameters for the wrong equation.</p><p>To bridge this gap, the machine learning world has revived and supercharged a technique called <strong>Symbolic Regression (SR)</strong>: it doesn’t solve for parameters; it solves for the law structure itself.</p><p>Unlike traditional regression that fits parameters to a pre-defined rigid structure, Symbolic Regression searches the infinite, open-ended space of mathematical operators (addition, multiplication, sine, exponentials, etc.) to build an equation’s structure from scratch. It doesn’t just find the weights; it discovers the formula itself.</p><p>Today, two prominent open-source projects — <em>PySR</em> and <em>AI Feynman</em> — illustrate the two main paradigms in this space. They represent two distinct approaches (evolutionary vs. physics-first). Let’s look at how they work and why they matter.</p><p>In the next sections, you’ll see two common ways SR is used in practice: <strong>to discover laws directly from experimental/field data</strong> and to <strong>distill a slow numerical simulators into a portable closed-form equation</strong> for fast decision-making.</p><h4><strong>PySR: Evolving the Complexity</strong></h4><p>Built by Miles Cranmer (<em>Princeton University</em> and <em>Flatiron Institute</em>), PySR is a high-performance library that uses Regularized Evolution. It doesn’t use neural networks internally; instead, it evolves equations like a biological species. It generates random formulas, competes them against each other (survival of the fittest), and mutates the survivors. It relies on a speed-optimized backend (written in Julia) to check millions of equations per second [1].</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fChYsc4gayAWl5oAF1wGrw.png" /><figcaption>PySR’s inner working. Source: <a href="https://github.com/MilesCranmer/PySR">https://github.com/MilesCranmer/PySR</a></figcaption></figure><p>One of the biggest challenges in Symbolic Regression is the research space. If you ask an algorithm to find an equation that perfectly fits 1&#39;000 noisy data points, it might generate a monstrous, 500-term polynomial that overfits the noise and completely misses the underlying physics.</p><p>PySR solves this brilliantly by allowing the data scientist to explicitly define and tune the <em>complexity</em> of the equations. Every mathematical operation has a cost. An addition symbol (`+`) might cost 1 point of complexity, while a sine wave (`sin`) might cost 2 points. As PySR searches for equations, it doesn’t just output one final answer. It presents you with a <strong>Pareto front, </strong>a set of solutions you choose to get the most efficient combinations, tipically used where improving one objective necessarily worsens another.</p><p>Think of the Pareto front as a shortlist of equations offering the best possible trade-offs between mathematical simplicity (low complexity score) and predictive accuracy (low loss). You can look at the menu and select the sweet spot equation that captures the physical dynamics without overfitting.</p><p>This complexity tuning has led to massive breakthroughs. In a landmark paper, astrophysicists used PySR to rediscover the laws of Dark Matter cosmology. They first trained a Graph Neural Network (GNN) to simulate how dark matter halos cluster together. However, the GNN was a black box. By feeding the GNN’s internal messages into PySR, the Symbolic Regression algorithm distilled the neural network’s complex vector math into a single, compact equation. It didn’t just fit the data; it actually rediscovered a variant of a known analytic formula that scientists had derived by hand decades earlier [2].</p><p><strong>Quick start: run PySR in ~20 lines (Python)</strong></p><pre># pip install pysr<br># First run will download/configure a Julia backend (it may take a few minutes).<br><br>import numpy as np<br>from pysr import PySRRegressor<br><br># 1) Create a tiny toy dataset (replace X, y with your real data)<br>rng = np.random.default_rng(0)<br>X = rng.normal(size=(300, 2))<br>y = 3.0 * X[:, 0] ** 2 + 2.0 * X[:, 1] + 0.1 * rng.normal(size=300)<br><br># 2) Define the symbolic search space (operators) + run the search<br>model = PySRRegressor(<br>    niterations=200,                  # increase for better results<br>    binary_operators=[&quot;+&quot;, &quot;-&quot;, &quot;*&quot;, &quot;/&quot;],<br>    unary_operators=[&quot;square&quot;, &quot;sin&quot;, &quot;cos&quot;, &quot;exp&quot;],<br>    maxsize=20,                       # complexity cap (helps fight bloat)<br>    model_selection=&quot;best&quot;,           # or &quot;accuracy&quot;/&quot;simplicity&quot;<br>)<br><br>model.fit(X, y)<br><br># 3) Inspect the Pareto front (accuracy vs complexity) and pick an equation<br>print(model.equations_)              # list of candidate equations<br>print(&quot;Best equation:&quot;, model.get_best())</pre><h4><strong>Genetic Algorithms as Optimizers</strong></h4><p>But how exactly does a tool like PySR search an infinite space of mathematical symbols to build that Pareto front without crashing your computer?</p><p>The answer lies in the <em>metaheuristic optimization algorithms</em>: a high-level computational strategy that leverages mathematical probability and precise set of rules to efficiently navigate complex multidimensional spaces.<br>PySR uses, for instance, <strong>evolutionary strategies</strong> to explore equation structures, while also optimizing numerical constants inside promising candidates. Imagine a <em>survival game</em> for equations: it generates random formulas, tests them against your data, and lets the best ones breed to create better formulas.</p><p>PySR pull elegant, physics-obeying equations out of complex datasets. But this raises a fundamental question: <em>How do you optimize an equation?</em></p><p>In a standard neural network, the architecture (the layers and nodes) is fixed, and you use gradient descent to smoothly update the continuous weights. But in Symbolic Regression, the <em>structure</em> of the equation itself is the variable. You cannot calculate the gradient of a plus sign (`+`) with respect to a sine wave (`sin`). The search space of all possible mathematical formulas is discrete, infinite, and non-differentiable.</p><p>To navigate this infinite space efficiently, Symbolic Regression relies on a biologically inspired optimization engine: <strong>the Genetic Algorithm (GA)</strong>.</p><p>To understand how equations evolve, you first have to stop thinking of equations as flat strings of text, and start thinking of them as <strong>Expression Trees</strong>. In an expression tree, the internal branches are mathematical operators (e.g., `+`, `*`, `sin`), and the outer leaves are the input variables (`x`, `y`) or numerical constants.</p><p>Here is how the evolutionary loop transforms a forest of random trees into a discovered law of physics:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UpdBz_Pl734gyIMgczf7Eg.png" /><figcaption>Functional scheme of Symbolic Regression. Image by the author.</figcaption></figure><p><strong>The Evolutionary Loop</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/828/1*X7fotHRn0Om84BKjz7c1Iw.png" /></figure><h3><strong>AI Feynman: a Framework to Mimick Human Genius</strong></h3><p>The second heavyweight is <strong>AI Feynman</strong>, developed by Max Tegmark’s research group at <em>MIT</em> [3].</p><p>The name is a deliberate homage to the legendary Nobel laureate physicist, Richard Feynman. Feynman was famous for his incredibly intuitive, almost playful approach to problem-solving. If a physics problem was too hard, Feynman didn’t just smash brute-force mathematics against it. He looked for symmetries, analyzed physical units, looked at the extreme limits of the problem (e.g., “what happens if the mass is zero or infinity?”). Then he would break a massive, seemingly unsolvable system into smaller, independent, solvable pieces — a <em>divide and conquer</em> strategy.</p><p>The AI Feynman framework was built to literally mimic Richard Feynman’s human intuition.<br>Instead of directly use evolutionary algorithms on raw data like PySR, AI Feynman uses deep neural networks as scouts. It first trains a Deep Neural Network (DNN) on the raw data, to approximate the unknown function. DNN acts as a differentiable, noiseless surrogate model. Then, the framework it interrogates its own neural network to hunt for Feynman-like properties: Does this data have translational symmetry? Is it rotationally invariant? Can this giant equation be mathematically separated into two smaller, independent equations?</p><p>Using the DNN surrogate, AI Feynman generates synthetic input grids (holding some variables fixed while sweeping others) to test properties like <strong>additive</strong> or <strong>multiplicative separability</strong>. When a separation hypothesis holds, it splits the target equation into smaller modules. Also, because the surrogate is made differentiable, it can use <strong>gradients / partial derivatives</strong> to detect structural patterns (e.g., dependence only on a difference or ratio of some features), revealing translational or scaling symmetries and collapsing variables into a single engineered feature. This divide-and-conquer loop repeats until each module is simple enough to solve with standard polynomial fits or a smaller Symbolic Regression.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-p4q2NnAD87LJQ-zuufatA.png" /><figcaption>Inner workflow of AI Feynman. Image by the author.</figcaption></figure><p>This physics-inspired strategy is invaluable for finding order in chaos. A recent extension of this architecture [4], was unleashed on the notoriously chaotic three-body problem or non-linear Schrodinger equation. Finding conserved quantities (like energy or momentum) in chaotic systems is generally difficult. Yet, by using this divide-and-conquer strategy on data, the AI acted like a physicist, uncovering hidden structural rules inside chaos.</p><p>These tools represent a paradigm shift. We are no longer just training models to blindly predict data; we are training them to speak the language of the universe.</p><p>To summarize, AI Feynman uses neural networks as scouts to simplify the data, looking for <strong>symmetries</strong> and <strong>separability</strong>. It then tears down the simplified pieces against a library of known mathematical forms using Symbolic Regression.</p><h3><strong>Synergies</strong></h3><p>We have explored three distinct pillars of Scientific Machine Learning (SciML):</p><p>- <strong>Inverse Problems:</strong> Forcing a neural network to play detective and hunt down hidden physical parameters from sparse observation data</p><p>- <strong>Surrogate Models:</strong> Trading a fraction of accuracy for a strong speed-up over classical physics simulators</p><p>- <strong>Equation Discovery:</strong> Evolving raw data into human-readable, mathematically valid equations using genetic algorithms</p><p>It is often easy to view these as isolated tools. But the true power of SciML is unlocked when you combine them into a single, cohesive workflow.</p><p>Let’s look at a practical, real-world example.<br>Imagine you work on a renewable energy project. Your first major task is to optimize the design of a novel, super-efficient wind turbine blade. The engineering team has an accurate and slow Computational Fluid Dynamics (CFD) simulator, but they only have sparse wind tunnel data for the new blade material.</p><p>Here is how you would deploy the SciML trinity to solve the problem:</p><p><strong>Phase 1: The Detective Work (Inverse Problem)</strong></p><p>You can’t simulate the aerodynamics of the new turbine blade if you don’t know the exact drag coefficient of the experimental blade. The engineers only gave you a few data points from a wind tunnel test.</p><p>You set up a Physics-Informed Neural Network (PINN). You feed it the sparse wind tunnel coordinates and embed the Navier-Stokes equations (fluid dynamic laws) directly into the loss function. You declare the unknown drag coefficient as a trainable parameter. As the network trains, it simultaneously learns to reconstruct the full airflow and mathematically deduces the drag coefficient of the new material.</p><p><strong>Phase 2: Cheating Time (The Surrogate Model)</strong></p><p>Now that you have the missing physical parameter, the engineering team wants to test 10&#39;000 different geometric variations of the blade design. Running 10&#39;000 simulations through their classical CFD solver would take months on HPC.</p><p>Instead, you use the classical solver to generate just 500 carefully selected high-fidelity simulations. You use those 500 runs to train an <strong>AI Surrogate</strong> (like a GP, CNN, GNN or Neural Operators). Your trade-off is to capture the underlying physical mapping of the airflow with just enough accuracy for the job. Once trained, you run the remaining 9&#39;500 variations through the surrogate. What would have taken months, now can takes you minutes.</p><p><strong>Phase 3: The Ultimate Deliverable (Equation Discovery)</strong></p><p>The surrogate model is incredibly fast, but it is still a <strong>black box</strong>: it gives you numbers, not a statement you can <em>inspect</em>. In many engineering settings, the <em>deliverable</em> isn’t just a fast predictor — it’s something you can <strong>verify, communicate, and reuse</strong>.</p><p>Now take the massive dataset generated by your surrogate model and feed it into PySR. The genetic algorithm spins up, mutating and mating thousands of mathematical expression trees. It hands you a Pareto front of possible equations, and you can select a highly accurate or elegantly simple equation that describes the pressure distribution over the new blade geometry.</p><p>We add S<strong>ymbolic Regression</strong> even when we already have a CFD simulator and a trained surrogate, for several reasons:</p><ul><li><strong>Portability: </strong>a compact equation can be shipped anywhere (spreadsheets/Excel, dashboards, optimization scripts) and evaluated in milliseconds, without GPUs, a Python stack, or a CFD environment.</li><li><strong>Interpretability + trust:</strong> an explicit formula is easier to sanity-check with subject-matter experts (limits, monotonicity, dimensional consistency, sensitivity).</li><li><strong>Distillation of complex solvers:</strong> numerical simulators already encodes known equations, but the results emerges only after heavy numerical discretization and boundary conditions. Symbolic Regression can <strong>compress</strong> that expensive numerical pipeline into an analytical closed-form approximation.</li><li><strong>Sometimes there is no simulator:</strong> in many real-world problems the data comes from experiments/field sensors, and the governing equation is unknown or only partially known. In that case SR isn’t compressing a solver — it is a tool for <strong>equation discovery</strong> directly from reality.</li></ul><p>In a this way, you didn’t just fit a curve. You discovered missing constants, bypassed a computational bottleneck, and derived a novel equation from scratch. Like an AI-augmented physicist.</p><p><strong>— — — — —</strong></p><p>The era of treating machine learning models as blind, data-hungry black boxes is ending. For scientists, SciML and Physics-Informed AI represents a profound paradigm shift.<br>By embedding known physical laws into our AI architectures, we aren’t just making our models more robust to out-of-distribution data. We are making them exponentially faster with surrogate models, capable of uncovering hidden parameters through inverse problem-solving, and elegantly explainable via Symbolic Regression. You are no longer just interpolating data points; you are modelling physical laws.</p><p><strong>— — — — —</strong></p><p><strong>References</strong></p><ul><li>[1] Cranmer, M., et al. — PySR architecture and Pareto front optimization</li><li>[2] Wong, K., et al. — Automated discovery of interpretable gravitational-wave population models</li><li>[3] Udrescu, S.-M., Tegmark, M. — AI Feynman: A physics-inspired method for symbolic regression</li><li>[4] Liu, Z., Tegmark, M. — AI Poincaré 2.0: Machine Learning Conservation Laws from Differential Equations</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d02e6aba13cf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/from-black-boxes-to-equations-the-practical-magic-of-equation-discovery-with-ai-d02e6aba13cf">From Black Boxes to Equations: The Practical Magic of Equation Discovery with AI</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When BERTopic Fails, Let  OpenAI GPT Summarization Fix Your Topics & Clustering]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/eni-digitalks/when-bertopic-fails-let-openai-gpt-summarization-fix-your-topics-clustering-570ac67ede0b?source=rss----fe373d427779---4"><img src="https://cdn-images-1.medium.com/max/756/1*pBGzDFrVXJxzLNq7vPGAjQ.png" width="756"></a></p><p class="medium-feed-snippet">This research focuses on a common business scenario where customer feedback is collected through surveys after a product or service is&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/eni-digitalks/when-bertopic-fails-let-openai-gpt-summarization-fix-your-topics-clustering-570ac67ede0b?source=rss----fe373d427779---4">Continue reading on Eni digiTALKS »</a></p></div>]]></description>
            <link>https://medium.com/eni-digitalks/when-bertopic-fails-let-openai-gpt-summarization-fix-your-topics-clustering-570ac67ede0b?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/570ac67ede0b</guid>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Maria Vallarelli]]></dc:creator>
            <pubDate>Wed, 01 Oct 2025 14:59:59 GMT</pubDate>
            <atom:updated>2025-10-01T14:59:35.925Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Crafting a successful loyalty program with Business Design and Data-driven approach]]></title>
            <link>https://medium.com/eni-digitalks/crafting-a-successful-loyalty-program-with-business-design-and-data-driven-approach-40597873d887?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/40597873d887</guid>
            <category><![CDATA[research]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[business-design]]></category>
            <category><![CDATA[loyalty]]></category>
            <dc:creator><![CDATA[Alicja Gackowska]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 14:42:11 GMT</pubDate>
            <atom:updated>2025-09-22T14:42:10.946Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/960/1*fivxUnjg7RMY9bLjlq3NjQ.png" /><figcaption>Source: <a href="https://unsplash.com/it">Unsplash</a></figcaption></figure><p><em>How Eni Plenitude’s loyalty program combines customer-centric strategies and rigorous testing to deliver value to both: customers and business?</em></p><p><em>This article was written with </em><a href="https://medium.com/@silvia.bajoni"><em>Silvia Bajoni</em></a><em>.</em></p><p>In today’s competitive market, creating innovative solutions that meet customer needs while generating value for the company is crucial. This article explores how Eni Plenitude designed and launched its customer loyalty program, “Plenitude Insieme,” and how a Business Design and Data-driven approach contributed significantly to its success. It highlights the benefits of applying these methodologies to create solutions that are meant to be equally <strong>desirable, viable </strong>and<strong> feasible</strong> and by identifying the riskiest assumptions in these three areas, test them, and gain confidence in the chosen path.</p><figure><img alt="Diagram with overlapping circles labeled Desirability, Viability, and Feasibility; the shared center is shaded." src="https://cdn-images-1.medium.com/max/1024/1*LT5iQOMZlgoXtTdSANCsNg.jpeg" /><figcaption>IDEO’s Desirability, Viability and Feasibility Framework &lt;source: <a href="https://makeiterate.com/ideos-desirability-viability-feasibility-framework-a-practical-guide/">https://makeiterate.com/ideos-desirability-viability-feasibility-framework-a-practical-guide/</a>&gt;</figcaption></figure><p><strong>Business Design: Merging Business and Design</strong></p><p>Business Design is an intersection of design methodologies, such as Design Thinking and Service Design, with business requirements to create solutions that are profitable and technically feasible. This approach is gaining recognition because not everything that people need or love can generate revenue and ROI (Return on Investment). Eni Plenitude’s design and business team faced the challenge of creating and securely launching its customer loyalty program using this methodology.</p><figure><img alt="Diagram with two overlapping circles: ‘Design Thinking &amp; Service Design’ linked to Creativity, and ‘Business Thinking’ linked to Structure. The overlap is labeled ‘Business Design’" src="https://cdn-images-1.medium.com/max/1024/1*lCqU9lCHX_oYfyz22BmK9A.png" /><figcaption><em>Business Design Scheme &lt;source: </em><a href="https://www.linkedin.com/pulse/what-business-design-try-business-design-7vu1f%3e"><em>https://www.linkedin.com/pulse/what-business-design-try-business-design-7vu1f&gt;</em></a></figcaption></figure><p><strong>How Business Design supported the program creation?</strong><br> <br> The project was developed by a cross-functional team of experts in marketing, data, branding, communication, market research, and design. They collaborated throughout the entire development process, leveraging diverse skills and expertise.</p><p>The Business Design methodology supported the program creation in three main aspects:</p><p><strong>1.</strong> <strong>Co-designing the program strategy</strong></p><p><strong>2.</strong> <strong>Applying customer and user-centric approach</strong></p><p><strong>3.</strong> <strong>Defining riskiest assumptions and testing them before implementation</strong></p><p>Let’s dive into these points in detail:</p><p><strong>1.</strong> <strong>Co-Designing the program strategy</strong></p><p>In typical corporate settings, developing this kind of cross-functional initiative meant to navigate high complexity and manage numerous stakeholders. To create a clear program strategy, business designers assisted the team in setting up and organizing various collaborative activities. These activities were essential for gathering fundamental information and establishing a unified vision for everyone in the company.<br> <br> The co-design process included several key activities:</p><p>a) <strong>Gathering market and customer data</strong>: The team collected all available in-house market and customer data to identify main insights, pain points, and opportunities from the customer’s perspective. This data was crucial in understanding customer needs and preferences, which informed the design of the loyalty program.</p><p>b) <strong>Conducting internal strategic interviews</strong>: To align the program vision with the company strategy, the team conducted internal strategic interviews with key stakeholders. These interviews helped set expectations, identify priority areas that could generate value, and gather relevant information about the program vision.</p><p>c) <strong>Organizing co-design workshops</strong>: The team organized co-creative workshops with key stakeholders to define the program’s strategy, goals, and unique value proposition. Tools such as Blue Ocean Mapping and Strategy Canvas were used to facilitate these workshops and ensure a clear and aligned strategy.</p><figure><img alt="People in a workshop placing sticky notes on large whiteboards with grid layouts; others engage in similar activities in the background." src="https://cdn-images-1.medium.com/max/1024/1*yCmENkWQAytO-hl6lMusgA.jpeg" /><figcaption><em>Co-design workshop with project team and various stakeholders</em></figcaption></figure><p><strong>2.</strong> <strong>Applying a customer and user-centric approach</strong></p><p>The design and development of the new loyalty program were driven by a customer and user-centric approach, ensuring that user needs were at the core of the process to deliver the best possible customer experience. This approach fostered a Test &amp; Learn culture, minimizing risks by enabling the team to identify and address issues early, accelerate improvement actions, and refine solutions based on user feedback.</p><p>Key activities in this approach included:</p><p>a) <strong>Using customer data</strong>: The team used both quantitative and qualitative customer data to inform the program strategy and design. This data-driven approach ensured that the program was tailored to meet customer needs and preferences.</p><p>b) <strong>Developing customer journeys</strong>: Tailored customer journeys were created for various segments to decide which touchpoints and interactions customers would have with the loyalty program. This helped in designing a seamless and engaging customer experience for different targets.</p><p>c) <strong>Testing concepts and user interfaces</strong>: Concepts and user interfaces were tested to gather feedback and make necessary adjustments. This iterative process ensured that the final solution was user-friendly and met customer expectations.</p><p><strong>3.</strong> <strong>Defining and testing riskiest assumptions</strong></p><p>Developing and launching an initiative like a customer loyalty program is related to a list of risks (and opportunities) for a company with several millions of customers. To minimize uncertainty, the team tested assumptions through experiments run during the entire design process and evaluated numerous hypotheses. Typically, it’s beneficial to start the project by verifying its business model, then move on to more detailed assumptions regarding the project’s value proposition, functionalities and even economics.</p><p>Examples of assumptions fall into known already three areas: desirability risks (do customers really want or need it?), viability risks (will we have enough customers to reach the goals?), and feasibility risks (are we able to develop and renew the offer?).</p><p>The experiment process involved the following steps:</p><p>a) <strong>Defining measurable hypotheses</strong>: The team defined measurable hypotheses that could be confirmed or rejected directly by customers. This involved identifying key assumptions and formulating them into testable statements.</p><p>b) <strong>Finding testing methods</strong>: The team identified the best methods to test the hypotheses quickly and with a limited budget. They conducted low-cost experiments to gain valuable insights without significant investment.</p><p>c) <strong>Setting metrics and confirmation levels</strong>: Metrics were defined to measure the success of the experiments, and minimum levels of confirmation were set to determine whether the hypotheses were supported or rejected. This helped in making data-driven decisions throughout the development process.</p><h3><strong>Data-driven approach: how to learn fast?</strong></h3><p>A Data-driven approach was adopted since the very beginning, through different customer research activities:</p><p><strong>1.</strong> <strong>Concept tests</strong></p><p><strong>2.</strong> <strong>Usability tests</strong></p><p><strong>3.</strong> <strong>Customer satisfaction monitoring</strong></p><p><strong>1.</strong> <strong>Concept tests</strong></p><p>First of all, what is a concept test? It is a type of customer research that enables to test new products or services before their launch on the market. The goal of this type of research is to assess consumers’ interest in a new solution, by analyzing its alignment with needs and desires, distinctiveness with respect to currently available alternatives, likes, dislikes and, of course, intention to buy. At an early stage, testing more than one concept can also be useful to identify the one with the highest potential.</p><p>The insights obtained through a concept test should then be used to decide whether to actually launch a new solution or not, since investing in poorly performing products or services can be a dangerously costly mistake. Moreover, in the former case, they can provide precious optimization suggestions to apply before proceeding with the launch, in order to maximize the probability of success.</p><p>Ideally, a blend of qualitative and quantitative customer research methodologies should be used when testing a new concept, since qualitative research allows to deeply explore consumer’s needs and expectations and gather ideas, while quantitative research enables to measure the potential of the solution.</p><p>As soon as the first draft of the Eni Plenitude’s loyalty program was ready, a concept test involving customers was organized. First of all, the team explored customers’ attitudes towards loyalty programs in general, in terms of perceived relevance, sought-after benefits, as well as actual subscription and participation. Then, the team tested different types of possible benefits and mechanics, to identify the most appreciated ones and to measure intention to subscribe to a loyalty program entailing such features offered by Eni Plenitude. Finally, the team investigated subscription and communication preferences.</p><p><strong>2.</strong> <strong>Usability tests</strong></p><p>The main goal of a usability test is to assess the effectiveness of new products or services by observing how consumers interact with them. During these tests, consumers are asked to perform a series of tasks related to a new solution, usually through specific interfaces, with the support of a researcher who acts both as a guide and an observer.</p><p>Usability tests enable acquiring familiarity with consumer behavior and can easily help identify faults or gather suggestions for improvement of a new solution. Once again, they can minimize the risks entailed by the launch of new products or services and raise their market potential.</p><p>One of the most common usability test techniques is “thinking aloud,” which requires consumers to think out loud while performing different tasks, verbally expressing their actions, intentions, and thoughts during the interaction with the solution in test.</p><p>As soon as the interface of the loyalty program platform was ready, a usability test with the thinking aloud technique was organized, involving Eni Plenitude customers. The test aimed to assess the clarity and completeness of the information present on the platform, its ease of navigation, and its look and feel. Moreover, the usability test was paired with an eye-tracking tool, able to record eye movements, allowing the collection of information about the interface elements best able to capture customers’ attention.</p><p>Additionally, the usability test of the new loyalty program was repeated a few months afterward, following an update of the platform, in a continuous learning-continuous improvement logic.</p><p><strong>3.</strong> <strong>Customer satisfaction monitoring</strong></p><p>Apart from the tests conducted before the launch of the loyalty program or its updates, its impact was then indirectly assessed through another research: customer satisfaction monitoring. This type of research is an ongoing quantitative study that involves a robust sample of Eni Plenitude customers, aimed at measuring the main KPIs of customer satisfaction (e.g. Net Promoter Score, declared loyalty, etc.) and observe how they evolve over time.</p><p>To measure the performance of the loyalty program, an oversample of its subscribers was integrated in the sample of the ongoing customer satisfaction monitoring. Then, some analyses were conducted to assess statistically significant differences in customer satisfaction KPIs between loyalty customers vs. Eni Plenitude customers overall. Moreover, loyalty customers involved in the research were further divided in two groups: active vs. passive participants, to capture differences also among these two types of customers.</p><h3><strong>Conclusions: benefits of the Business Design and Data-driven approach</strong></h3><p>The Business Design and Data-driven approach used by Eni Plenitude in developing its loyalty program highlights the importance of integrating customer insights, collaborative strategy development, and iterative testing. Throughout the process, Business Design methodologies were coupled with customer research to validate the loyalty program with Eni Plenitude customers before its market launch. By fostering a Test &amp; Learn culture and continuously monitoring performance, Eni Plenitude created a loyalty program that effectively engages 90% of its subscribers, significantly surpassing the market benchmark of 40%. This approach minimized risks and ensured the program met customer needs while delivering value to the company, ultimately contributing to its success.</p><h3><strong>Annotation: Eni Plenitude’s loyalty program</strong></h3><p>The Eni Plenitude’s loyalty program, launched in Italy in December 2022, rewards customers for their loyalty and engages them in energy-related and sustainability topics. By the end of 2024, the program had reached 1 million users with an interaction rate exceeding 90%, establishing itself as a comprehensive ecosystem of benefits, experiences, and opportunities. It aims to deepen the connection between the company and its customers by creating a new alliance that goes beyond the transactional relationship of payment and sale. It aspires to establish a lasting bond with customers based on the value created together over time and on the sharing of social and environmental values particularly important to Eni Plenitude.</p><p>In 2024, the “Plenitude Insieme” program received the <em>Best Platform</em> award and a special mention for the <em>“Rewards Catalogue”</em> at the <em>Promotion Awards 2024</em>, a key event in the Italian relational marketing landscape for over thirty years, organized by Promotion Magazine, which covers loyalty processes, customer experience and data-driven marketing. Additionally, the program is a finalist in the <em>“Best Loyalty Program Utilities”</em> category at the <em>2025 International Loyalty Awards</em>. These recognitions highlight the success of “Plenitude Insieme” in transforming a traditional loyalty program into an engaging and rewarding experience.</p><p>For more information, visit the “Plenitude Insieme”: <a href="https://insieme.eniplenitude.com/">https://insieme.eniplenitude.com/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40597873d887" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/crafting-a-successful-loyalty-program-with-business-design-and-data-driven-approach-40597873d887">Crafting a successful loyalty program with Business Design and Data-driven approach</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Innovation from within | Is it possible to find good ideas at scale?]]></title>
            <link>https://medium.com/eni-digitalks/innovation-from-within-is-it-possible-to-find-good-ideas-at-scale-4adc3d2780b7?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/4adc3d2780b7</guid>
            <category><![CDATA[service-design]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[business-design]]></category>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[energy]]></category>
            <dc:creator><![CDATA[Marina Po]]></dc:creator>
            <pubDate>Mon, 14 Jul 2025 16:15:20 GMT</pubDate>
            <atom:updated>2025-07-14T16:15:20.623Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gAbF81DW3d7HUhNfGn8rxg.png" /><figcaption>AdobeStock_1309044086 by Alex</figcaption></figure><p><em>How can teams innovate within the company in a genuine way? A look inside to make a step outside the box and move towards transformative digital growth that enhances company’s productivity.</em></p><p>If you put 12 people in a room from various company units and ask them to list all the problems they are facing, probably under within an hour you would have a pretty long list to go thru. Experienced workers can quickly identify most of the operational, logistical, and administrative challenges of a company. While this list provides valuable insights, it’s essential to be able to identify the most critical issues with reasonable efforts. More importantly, can we think of any good ideas that can address these problems? And will our proposed solutions yield wanted results?</p><p>In this article I will dive deep into the mechanisms that incite innovation. Keep reading if you are someone that’s trying to innovate within a company be it by transforming mindset or by creating and implementing new products.</p><p><strong>Instead of creating long lists can we do it better?</strong></p><p>Yes, thankfully by adopting systemic and human-centric approaches for solving problems, we can efficiently extract insight in reasonable time from a minimum sample needed. Human-centric approach campaigns that we should always keep in mind what the person wants and needs. We should continually ask ourselves: ‘What do our users, including internal employees, require to improve their performance and better serve our clients and stakeholders?’</p><p>System thinking teaches us to view problems holistically, rather than focusing solely on symptom analysis. By looking for root causes and understanding how problems are interconnected, we can develop more effective solutions. This approach is often referred to as ‘seeing the bigger picture.’ For example, when addressing warehouse overflow, system thinking might involve examining factors like shipping documents, work orders documents, context of warehouse workers, rather than simply focusing on building bigger warehouses.</p><p><strong>The approach: Creating space for conversations</strong></p><p>Design is renowned for bringing people together on account of its principles collaboration, and co-design. The seamless collaboration of sharing ideas and opinions no matter the ranking and hierarchy. What design thinking method does well it uses simple tools to process divergent and convergent thinking. This happens particularly well in a workshop-like circumstances.</p><p>So, yes put 12 people in a room but in a designerly manner. Don’t make them do lists of problems but make them think about root causes and opportunities. Use affinity mapping to declutter and decompose complicated problems. The ‘needs’ should become challenges to solve and an anchor for opportunities. Most of the times if problems are not shared in systemic manner, it’s hard for the group to understand where their efforts should be focused.</p><p><strong>How Eni’s operational business put it in practice?</strong></p><p>When one of the Eni’s operational businesses wanted to drive innovation on the digital front, they were not sure how to approach it inclusively at first. That’s how a collaboration between digital &amp; tech partners and the design team came to be. Together, we identified 4 areas of interest to explore: production, production management, maintenance, and warehouse management. The aim was to bring together small groups of very busy people and give them time and place to converse about their problems and opportunities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1004/1*3YMKh74fTJjF9Nz8tVCG9w.png" /><figcaption>Scheme of 5 steps — Design approach for innovation</figcaption></figure><p>Here are the 5 phases of the approach we adopted and executed:</p><p>1. <strong><em>Discovery</em></strong> — Preliminary research to gather additional qualitative data.</p><p>2. <strong><em>Opportunity</em></strong> — Starting from the right problems we can identify innovation opportunities.</p><p>3. <strong><em>Design</em></strong> — Co-design with the users themselves.</p><p>4. <strong><em>Prototype</em></strong> — Try it out, see what works better and how it feels to the users.</p><p>5. <strong><em>Test &amp; Validation </em></strong>— Test it and validate it with users and stakeholders.</p><p>During the discovery phase, we interviewed people from various business units to map out their daily tasks, gaining a broader perspective on responsibilities, processes, and more. In the opportunity phase, we employed participatory design techniques. We gathered individuals from the four areas of interest in a room to work together in groups and discuss challenges based on journey maps derived from the interviews. The working group created idea cards and collaboratively selected the most promising ones. Subsequently, tech experts, UX designers, and digital professionals joined the groups to co-design a potential digital product. Together, they envisioned a future journey for the product, detailed its features, and designed the desired user experience. The team then proceeded to create a low-code or no-code prototype to test functionalities and user experience before industrializing the product.</p><p><strong>Conditions for a good innovation — What makes a difference?</strong></p><p>While the design thinking approach can help solve many problems, blindly applying it without considering specific conditions may limit the discovery of breakthrough ideas. To foster true innovation, it’s essential to account for the conditions and environment needed to support and inspire creative thinking.</p><p>In every step of the process, you should keep in mind the following questions:</p><ul><li>Is this the right problem we should be solving?</li><li>Is this idea considering the user needs?</li><li>Is this idea beneficial for the business overall?</li><li>Can we scale up the design?</li></ul><p>Moving to action, we need to implement tangible conditions that enable innovation at every stage. Start by creating a clear vision and well-defined goals. This vision should guide the entire innovation process, ensuring that everyone is aligned and working towards the same objectives. Continuous learning is pivotal. The employees need to stay up-to-date with the latest trends and technologies and keep their minds open to new possibilities. Additionally, collaboration across different departments and teams is key to encourage the sharing of ideas and knowledge. This also means being open to changing processes when we need to adapt to new conditions. Lastly, diffused ownership is a way of advocating for new ideas, garnering support, and driving change. Everyone needs to feel they can propose and discuss an idea.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1004/1*5MrgW7TQubdBLW3c0uN7yQ.png" /><figcaption>Scheme of 5 steps — Conditions for good innovation</figcaption></figure><p>The requirements for an advantageous innovation in a company environment:</p><ol><li>Having an open mind and creative attitude.</li><li>Having introduced the variety and the right skillset in the group.</li><li>Awareness and know-how of technological potential.</li><li>Build the right partnerships within and outside the company when needed.</li><li>Being timely, in decisions and concept proofing.</li><li>Having set the right conditions for collaboration.</li><li>Never exclude the user (employees) out of the conversation.</li><li>Being able to showcase your findings and ideas to stakeholders and broader users.</li><li>Engagement in advocating our solution based on strong evidence and support of our stakeholders and users.</li></ol><p>To generate truly innovative ideas, it’s essential to embrace diversity of thought. It’s not enough to bring people together. While collaboration among like-minded individuals that also have similar responsibilities can be productive, incorporating a tech or digital expert can elevate the process. These figures, actively engaged in discussions and ideation sessions, can provide a unique perspective, challenging assumptions and offering novelty previously unknown to the ‘base’ group.</p><p>Here are additional questions to always keep in mind:</p><ul><li><strong>Are there any partnerships we should be building?</strong></li><li><strong>Are we bringing new ideas in? Are we engaging with tech experts?</strong></li></ul><p><strong>What is a good idea?</strong></p><p>Now, that you have your ideas, you must choose the best ones. The best ideas should usually adhere to these 3 criteria:</p><p><strong>· useful for the users</strong></p><p><strong>· feasible for the business</strong></p><p><strong>· and have measurable benefits in terms of cost reduction or productivity enhancement</strong></p><p>To find a “good idea” you should avoid competition-like-decision-making that is based on some sort of rigid metric. It is better to opt for few ideas that can be scaled according to the available budget. A good idea represents a valuable opportunity for the business.</p><blockquote><strong><em>“What is a good digital opportunity?</em></strong></blockquote><blockquote><strong><em>When an idea addresses a genuine need or problem that can be leveraged by digital technologies to a scalable and impactful solution, and it demonstrates potential for both business value and user and/or employee experience.”</em></strong></blockquote><p><strong>Concept proofing: How Eni’s operational business choose the most viable ideas?</strong></p><p>Finally, after thorough looking and analysis our working group introduced to the stakeholders four ideas from four different areas of interest, each worth pursuing. Through strong collaboration with digital, tech and design experts, leveraging their extensive technology knowledge, and building and testing the prototypes, the business gained a clear understanding of potential next steps. Surprisingly, not all ideas needed to be developed from scratch. Some could evolve from existing products, others could be customized from market solutions and some would need to be custom-built.</p><p>The concept proofing process is essential for obtaining measurable insights into the advantages of your idea. It helps you advocate for your idea with management, wider user base, and partners. Ideally, this should be done by someone who is deeply familiar with the product, understands its design, its advantages. In our E2E products, the advocates are the employees themselves who envisioned the product.</p><p>In the last phases of our process, we should also do the following checks:</p><ul><li><strong>Are we considering the bigger picture?</strong></li><li><strong>Did our stakeholders understand the value?</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bHXVIpAExkidZoQ-Ry4qbA.jpeg" /><figcaption>Eni digiTALKS</figcaption></figure><p><strong>Conclusion</strong></p><p>Teams that work closely together often lack a broader perspective, focusing primarily on their immediate tasks and operational challenges. Systemic design approach can help address this by taking a holistic approach, examining businesses, teams, and problems within the context of a larger system. This broader view enables identification of systemic issues and opportunities for innovation.</p><p>As designers, our goal should be to guide the team toward a near-to-flawless user experience journey for a solution to be, regardless the technology. The priority should be on addressing user needs, helping the client focus on real business issues, and distinguishing between process-related and tool-related problems.</p><p>Often, issues arise when digital tools are outdated, overly complex or burdened by significant design debt, forcing teams to develop time-consuming workarounds. Another common issue is the lack of digital tools to automate tasks such as combining and handling large datasets. Teams frequently develop impressive skills in manually managing and manipulating vast amounts of information. While this can be admirable, it can also lead to a situation where these manual processes become normalized, hindering innovation and efficiency. By questioning the status quo and exploring potential automation and innovation opportunities, teams can free up time and resources for more strategic work.</p><p>Therefore, yes innovation can come within! It greatly helps if we cultivate cross-collaboration with new tech &amp; digital experts to harvest the potential of new technologies. Keeping an open mind and a big-picture vision in mind is where we should start. Finding the right problems to solve first and ensuring the business, the final user, is always included in product design discussions is fundamental. Lastly but of equal importance, showcasing the value and advocating the use cases to stakeholders is crucial for gaining support, feedback, and accelerating adoption!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4adc3d2780b7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/innovation-from-within-is-it-possible-to-find-good-ideas-at-scale-4adc3d2780b7">Innovation from within | Is it possible to find good ideas at scale?</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Experimental Design in the AI Era]]></title>
            <link>https://medium.com/eni-digitalks/experimental-design-in-the-ai-era-98f7cb095635?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/98f7cb095635</guid>
            <category><![CDATA[bayesian-optimization]]></category>
            <category><![CDATA[design-of-experiments]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[research-and-development]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Roberto S. Pinna]]></dc:creator>
            <pubDate>Wed, 26 Mar 2025 20:13:39 GMT</pubDate>
            <atom:updated>2025-03-26T20:13:39.755Z</atom:updated>
            <content:encoded><![CDATA[<h4>In this article, we will discuss how to construct this framework with AI, highlighting its primary benefits and addressing the key challenges involved in designing robust experiments.</h4><p>Article written in collaboration with <a href="https://medium.com/u/222a1d95645b">Federico Venturini</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*GXqoGP8iF0MwPhM5lVR1ew.png" /><figcaption>Credits: Image taken from Alamy.it</figcaption></figure><p>Experimental Design (ED) methods are extensively applied during the R&amp;D experimentation, focusing on identifying optimal experimental conditions and configurations to meet specific objectives. These methods aim to minimize the number of experiments required while maximizing the informational yield from the data collected.</p><p>Optimal Experimental Design (<strong>OED</strong>) is a modern approach within the ED framework, typically formulated as a Bayesian Optimization (<strong>BO</strong>) problem where a Machine Learning (<strong>ML</strong>) model is trained to estimate data, its uncertainty and find hidden knowledge in the data. It uses an explicit Acquisition Function (<strong>AF</strong>), which encompasses the target performance, uncertainty and the constraints of the problem. AF may incorporate criteria such as precision of parameter estimates, empirical error of model predictions or cost of the experiment.</p><p>OED presumes that some reliable experimental/synthetic data are already available. Thus, experimental results can be inferred with a surrogate function or model, cheaper to evaluate with respect to the physical collection of data or numerical simulations. This proxy can be extensively sampled and the search for the best parameters is typically performed with a gradient-free approach on the acquisition function.</p><p>The OED approach usually has 3 key elements:</p><ul><li>Definition of experimental domain</li><li>Building proxy model (surrogate) to approximate a black-box function (i.e. experiments)</li><li>Bayesian optimization of the target: An Acquisition Function and optimization to find which point to query for the next evaluationevaluation.</li></ul><p>Here is a typical OED loop that will be detailed in this article:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/742/1*085WxQsFUvmseKBDrhMsng.png" /><figcaption>OED loop workflow for typical applications</figcaption></figure><blockquote>In this article we will present the workflow for OED, the most used frameworks in python and an example use-case to clarify the concepts.</blockquote><h3>Dealing with Categorical Data</h3><p>The data type chosen to represent test design factors (inputs) and responses (outputs) in an experiment can have a major effect on the resources needed to conduct an experiment and the quality of its respective analysis.</p><p>Data can be classified as either numerical (quantitative) or categorical (qualitative).</p><p>Categorical data types are frequently encountered in experimental databases and represent a challenge on their own. In fact, this kind of data contains a relatively poor amount of information in comparison to numerical data types and often they are not natively handled by ML algorithms.</p><p>Using a numerical data type, especially a continuous one, maximizes test effectiveness, because test measurements require a much smaller sample size and produce tighter confidence intervals. It has been proved that using a binary response rather than a continuous one is equivalent to discarding from 38% to 60% of the test runs [<em>cohen1983</em>][<em>hamada2002</em>]. Also, if specification limits are provided, a numerical response can easily be converted to a categorical one, however it is not possible to do the opposite because of information loss. Nevertheless, most machine learning algorithms require input variables to be numeric.</p><p>There will be cases where you must deal with categorical data, thus you can choose different ML approaches such as integer encoding, one-hot encoding, leave-one-out etc. A common approach used while working with laboratory experiments, is <strong>Physics based encoding</strong>.</p><p>While traditional methods listed above are widely used, there are advantages in exploring approaches inspired by the physics underlying the process at study. As Data Scientist, is of outmost importance to deep dive into the specific domain knowledge and to talk with domain experts about the technologies at hand. Physics-based approaches aim to incorporate domain knowledge or physical principles into the encoding process. These methods are not usually quick and easy, but offer interesting possibilities:</p><ul><li>Capture domain expert knowledge to represent categorical variables and leverage known interactions with other quantities (e.g., swapping the name of a chemical substance with its measured properties).</li><li>Impose physical constraints on the encoded representations (e.g., you might enforce that certain categories cannot coexist simultaneously due to physical limitations). By incorporating such constraints, you ensure that the encoded features align with the underlying physics of the system.</li><li>Incorporate physics-inspired rules in the optimization process (e.g., conservation laws).</li></ul><p>Now that we got this point set up, we should move to the world of OED.</p><h3>Definition of the Experimental Domain</h3><p>In the field of OED, the experimental domain refers to the range of parameters (operative conditions) over which experiments are conducted.</p><p>It is of paramount importance to talk with the domain experts about the reasonable minimum and maximum limits of each experimental parameter: this is where the experimental activity will be focused.</p><h4>Domain mapping</h4><p>If the experimental activity is starting from scratch, it is probably useful to define the initial set of tests to start the exploration. In a multivariate context, there are some approaches that help to set up the most effective way to start exploring the domain, such as:</p><ul><li><strong>One Factor at a Time (OFAT</strong>): is a method in which the impact of change in one factor is studied on the output when all the other factors are kept constant. It’s the most basic mapping method and also the most inefficient one.</li><li><strong>Factorial methods</strong>: classical DOE methods where the impact of factors is studied by changing more factors at the same time. Factorial design is a powerful method used to study the effects of multiple variables on a single response, combining different values of the variables (called factors) to create many possible combinations. This kind of domain exploration can be performed for every combination (full factorial) or for just a subset of combinations (fractional factorial) to reduce the use of resources.</li><li><strong>Latin hypercube sampling (LHS)</strong>: the concept of LHS is well suited for multidimensional spaces and is based on the Latin square design. In a 2D Latin square, there is only one sample in each row and each column. A Latin hypercube generalizes this idea to an arbitrary number of dimensions. When using LHS, we divide the range of each variable (dimension) into equally spaced intervals. Sample points are then placed to satisfy the Latin hypercube requirements of orthogonality and to minimize correlation between different points. It’s important to consider that the number of divisions along a dimension axis remains equal for each variable.</li></ul><p>One of the main advantages of LHS is that it doesn’t require more samples as the number of dimensions increases: this independence from dimensionality is valuable.</p><h4>Data harvesting from literature</h4><p>It is advisable to make extensive research of the available data in the scientific literature prior the start of any experimental activity. The data gathered from literature can be precious and, when validated by the business, can be merged with current data to enrich the available dataset. Sources of scientific papers and referenced technical knowledge can be:</p><p><a href="https://scholar.google.com/">https://scholar.google.com/</a></p><p><a href="https://www.semanticscholar.org/">https://www.semanticscholar.org/</a></p><p><a href="https://www.sciencedirect.com/">https://www.sciencedirect.com/</a></p><p><a href="https://arxiv.org/">https://arxiv.org/</a></p><p><a href="https://www.perplexity.ai/">https://www.perplexity.ai/</a></p><h3>Building a proxy model to approximate a black-box function</h3><p>In the real world, experiments are usually like a black-box function where each query is slow and expensive. To tackle this problem, one should build a <em>proxy</em> model that resembles the output of the original black-box function and returns the related uncertainty quantification. This proxy must be faster and cheaper to query.</p><p>Below we list the most common models used to build a proxy for OED.</p><h4><strong>Gaussian Processes (GPs)</strong></h4><p>The most popular surrogate model in Bayesian optimization due to their flexibility and ability to provide a measure of uncertainty about the predictions. It provides probabilistic predictions with a mean and a variance, making them suitable for balancing exploration and exploitation. Gaussian Processes [<em>gp</em>] can be computationally expensive for large datasets because the complexity scales cubically with the number of data points. One can use GPs calling stand-alone APIs or can exploit BoTorch [<em>BT</em>], an integrated framework that encompasses GPs and Bayesian optimization.</p><h4><strong>Tree based models</strong></h4><p>Decision trees models provide a robust and non-parametric approach to model the objective function. They are less sensitive than GPs to the choice of hyperparameters and can handle high-dimensional data well. On the other hand, trees <em>do not naturally</em> provide a measure of uncertainty, but this can be approximated using methods like quantile regression forests or catboost. [<em>stat1</em>]</p><h4><strong>Bayesian Neural Networks (BNNs)</strong></h4><p>BNNs are neural networks with a Bayesian framework to model uncertainty in their predictions. They can capture complex patterns in data and scale well with large datasets. They are typically more challenging to train and require significant computational resources compared to GPs and tree based models.</p><h3>Sensitivity study</h3><p>Feature importance and sensitivity analysis are a crucial concept in machine learning that helps to understand the relative significance and impact of input features on the model prediction. Generally, these techniques assign a score to each input features based on how relevant they are for the model and help quantify how feature variations affect the output value. For example, models like XGBoost or random forest in general provide the access to the feature importance directly from the APIs after the model training.</p><p>There are also dedicated tools that exploit statistical methods for sensitivity analysis:</p><ul><li><strong>Permutation Feature Importance</strong>: it involves randomly shuffling the values of a feature and measuring its impact on model performance. The drop in performance indicates the feature’s importance (strong association with the target variable).</li><li><strong>Shapley method</strong>: It originates from cooperative game theory and provides a way to allocate the contribution of each feature to the overall prediction. Shapley values represent the average marginal contribution of a feature across all possible feature subsets. They ensure fairness by considering all possible combinations of features and their contributions. The Python library provides tools for calculating Shapley values, it gives unified framework for feature importance across different algorithms. [<em>shap1</em>]</li></ul><p>Sensitivity can be calculated for both regression (predicting numerical values) and classification (predicting class labels) problems. As a general rule, one should always perform a sensitivity study on the model, to better explain the output and to identify which features are less useful to the model. In fact, reducing the input feature by focusing only on the important ones can improve model efficiency and reduce the complexity of the OED.</p><h3>Bayesian Optimization of the target</h3><p>The next step is to build an acquisition function to be optimized. One of the most powerful tools in this regard is Bayesian Optimization (BO), which optimizes a function knowing the uncertainty of each measurement. For example, in the experimental case, it optimizes an expensive black-box system of the form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/226/1*bVm_0YPvdBJMIg568O3sRg.png" /></figure><p>where 𝒳 is the design space, and the objective function <em>f(x)</em> does not have a closed-form representation, does not provide function derivatives, and only allows point-wise slow and expensive evaluation through physical experiment. However, the evaluation of the function f could be corrupted by noise. Thus, the output we receive at point is</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/185/1*UnwMcwSdaFe9Yi-9XqP6Fg.png" /></figure><p>where the noisy term <strong>ε </strong>is usually assumed to be normally distributed with mean 0 and variance σ².</p><p><strong><em>In OED the black-box function f(x) is replaced by the cheaper proxy model discussed previously.</em></strong></p><p>To guide the optimization process, Bayesian optimization uses acquisition functions. The acquisition function is typically a computationally inexpensive function, evaluating how desirable a query point <strong><em>x</em></strong> is for the optimization problem. In other words, it uses the proxy model approach explained above, and it replaces our original problem with another optimization problem, but on a cheaper function. Each acquisition function can tune the tradeoff between <strong>exploration</strong> (searching new regions of the space) and <strong>exploitation</strong> (focusing on regions known to perform well).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HEcQwlBpnjwGbSzLgnQ0dg.gif" /><figcaption>Target function approximation by consecutive measurements suggested by the maximization of the acquisition function at each step. Source [baratchi2024]</figcaption></figure><p>Common acquisition functions include:</p><h4><strong>Upper Confidence Bound (UCB</strong>)</h4><p>It is the simplest acquisition function and contains explicitly the expected target value μ and its uncertainty σ at the domain point x, making a weighted sum of the two quantities:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/306/1*TTYGgdO3VK2f8Kq3LK5AMw.png" /></figure><p>The weight <strong>k</strong> is decided arbitrarily depending on the use case and this function allows for a straightforward tuning of the exploration-exploitation tradeoff.</p><h4><strong>Expected Improvement (EI)</strong></h4><p>Selects experiments that balance exploration (trying new configurations) and exploitation (using known information). It considers the difference between the current best value and the expected value of a new experiment. More exploratory, seeking regions where improvements are more likely. [<em>EIF</em>]</p><h4><strong>Probability of Improvement (PI)</strong></h4><p>Focuses on the probability that a new experiment f(x) will get an improvement I(x) upon the current best value.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/404/1*HIqIQRP7N2k3OJx448yNuQ.png" /></figure><p>It especially considers the cumulative distribution function (CDF) of the model improvement <strong>I</strong>. It is purely exploitative, favoring regions where is higher the chance of having high target values<em>.</em></p><blockquote>Each surrogate model has its strengths and weaknesses, and the choice often depends on the specific characteristics of the problem at hand, such as the dimensionality of the input space, the nature of the objective function, and the computational resources available. For more technical details about these acquisition functions, see [ACQF]</blockquote><h3>Customizing acquisition function</h3><p>After having chosen a proper acquisition function, it’s likely that customization can be required to fit the real use case needs. The experimental constraints that insist on possible designs can be numerous. Three categories of vincula are normally distinguished: <strong>operational constraints</strong>, <strong>weak constraints,</strong> and <strong>strong constraints</strong>. The same constraint can be seen as weak in one scenario and strong in another. ODE must take into account these constraints by embedding mathematical boundaries and penalties into the acquisition function.</p><p>It’s important to note that the vincula should be defined a priori to limit the exploration space.</p><h4>Operational constraints (a.k.a. boundaries)</h4><p>The classical formulation of the Design of Experiment assumes that the domain of operation of the process is an hypercube, but this space in the real world is not infinite and must obey to technical limits that cannot be overcome.</p><p>In fact, many situations could present the following problems:</p><ul><li>Some configurations may not be operable</li><li>Some inputs may be constrained by others</li></ul><p>In those cases, a proper customization of the acquisition function is required.</p><p>Examples of constraints you can encounter are:</p><ul><li>Configurations not feasible (or not relevant)</li><li>Experimental timescales that may collide with actual deadlines</li><li>Possibility or necessity to perform experiments in a particular order</li></ul><h4>Weak and strong constraints (a.k.a. penalties)</h4><p>Weak and strong constraints represent limitations that can be technically overcome but it is not convenient to do so. There is not a general definition of which constraints are weak or strong, these must be defined for each single case, together with the domain experts and in compliance with the business needs.</p><p>One basic example is the experimental cost, since the classical formulation of the Design of Experiment assumes that the cost of any configuration is equivalent. However, in reality the cost varies dramatically with respect to the configuration. Depending on budget availability, in some cases the cost can be a strong constraint while for others can be a weak one.</p><blockquote>This analysis (operational constraints + weak/strong constraints) should lead to a detailed formulation of the problem at hand. Only after this analysis, one can choose the correct approach. The constraint values must be defined beforehand.</blockquote><h3>Multi-objective problem</h3><p>In some use cases the design will involve multiple black-box targets to be optimized at the same time. When all the targets have to be maximized/minimized as much as possible, the correct approach is to include each target into the acquisition function as a new encompassing metric. One must choose appropriate weights, jointly with the domain experts, in order to normalize all the targets to the same order of magnitude and/or to incorporate the relative importance of each target into the acquisition function’s output. This step will boil-out to a single optimization problem.</p><p>For example, one simple approach is to perform a weighted sum of the targets to obtain a single metric <em>f*</em> :</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/295/1*DJdS5MFw8UPzbUK3ve2yOA.png" /></figure><p>In addition, when the constraints themselves are not known precisely, the global maximum of the acquisition function <em>f(x)</em> over some bounded domain is typically searched in <em>X </em>subject to the non-positivity of a set of constraint functions c₁, … , cₖ .</p><h3>Frameworks</h3><p>This section presents a series of libraries that can be used respectively for the definition of the experimental domain, for the construction of the proxy model and as a general framework for the Bayesian optimization.</p><p><a href="https://pythonhosted.org/pyDOE/)"><strong>PyDOE</strong></a><strong>: </strong>package designed to construct appropriate experimental designs. The package currently includes functions for creating designs for any number of factors. [<em>pydoe</em>]</p><p><a href="https://botorch.org/"><strong>BoTorch</strong></a><strong>:</strong> library for Bayesian Optimization research built on top of PyTorch. It provides a modular and easily extensible interface for composing Bayesian Optimization primitives, including probabilistic models, acquisition functions, and optimizers. [<em>BT</em>]</p><h3>Toy example</h3><p>To understand better this complex topic, let’s Imagine to be in a chemistry laboratory team which have to select the best recipe to create a lighter bioplastic, so the target would be to minimize the density of the final product, all other characteristics of the plastic material being equal. The variables that can be manipulated are the process temperature, the additive, and its concentration.</p><p>Before setting up any experiment would be wise to scan delve into literature for any available data on bioplastics and innovative production techniques, then put all the relevant data in a database.</p><p>Then experimental boundaries should be set on temperature, choice of the additive and its concentration. The name of the additives should possibly be substituted with their relevant chemical properties in order to work only with numerical features and improve the informational content.</p><p>One can use a mapping method to find the first optimal set of variables or use the literature database to build a proxy model, train it and then perform Bayesian optimization to find the most informative experimental variables. These suggested variables should guide the next experiments choice.</p><p>Finally, the bioplastic density resulting from each experimental choice will be added to the database in order to retrain the model and continue the loop.</p><h3>References</h3><ul><li>[cohen1983] Cohen J. et al., The Cost of Dichotomization. (1983)</li><li>[hamada2002] Hamada M. et al., The Advantages of Continuous Measurements over Pass/Fail Data. (2002)</li><li>[baratchi2024] Baratchi et al., Automated machine learning: past, present and future. (2024)</li><li>[zhang2023] Zhang M. et al., Bayesian Optimisation for Sequential Experimental Design with Applications in Additive Manufacturing. (2023)</li><li>[forrester2008] Forrester, A. et al., Engineering Design via Surrogate Modelling: A Practical Guide. (2008)</li><li>[greenhill2020] Greenhill, S. et al., Bayesian optimization for adaptive experimental design: A review. (2020)</li><li>[jones1998] Jones, D. R. et al., Efficient global optimization of expensive black-box functions. (1998)</li><li>[gp] <a href="https://scikit-learn.org/stable/modules/gaussian_process.html">https://scikit-learn.org/stable/modules/gaussian_process.html</a></li><li>[cat] <a href="https://catboost.ai/en/docs/">https://catboost.ai/en/docs/</a></li><li>[latin]<a href="https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.qmc.LatinHypercube.html">https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.qmc.LatinHypercube.html</a></li><li>[stat1] <a href="https://towardsdatascience.com/another-conformal-way-to-predict-probability-distributions-fcc63e78680d">https://towardsdatascience.com/another-conformal-way-to-predict-probability-distributions-fcc63e78680d</a></li><li>[shap1] <a href="https://www.analyticsvidhya.com/blog/2019/11/shapley-value-machine-learning-interpretability-game-theory/">https://www.analyticsvidhya.com/blog/2019/11/shapley-value-machine-learning-interpretability-game-theory/</a></li><li>[EIF]<a href="https://proceedings.neurips.cc/paper_files/paper/2017/file/b19aa25ff58940d974234b48391b9549-Paper.pdf">https://proceedings.neurips.cc/paper_files/paper/2017/file/b19aa25ff58940d974234b48391b9549-Paper.pdf</a></li><li>[ACQF]<a href="https://ekamperi.github.io/machine%20learning/2021/06/11/acquisition-functions.html">https://ekamperi.github.io/machine%20learning/2021/06/11/acquisition-functions.html</a></li><li>[BT] <a href="https://botorch.org/">https://botorch.org/</a></li><li>[pydoe] <a href="https://pythonhosted.org/pyDOE/">https://pythonhosted.org/pyDOE/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98f7cb095635" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/experimental-design-in-the-ai-era-98f7cb095635">Experimental Design in the AI Era</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Transforming Voice of Customer into Actionable Insights with BERTopic and OpenAI on Databricks]]></title>
            <link>https://medium.com/eni-digitalks/transforming-voice-of-customer-into-actionable-insights-with-bertopic-and-openai-on-databricks-aff3d4679011?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/aff3d4679011</guid>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[databricks]]></category>
            <dc:creator><![CDATA[Maria Vallarelli]]></dc:creator>
            <pubDate>Tue, 18 Feb 2025 17:04:55 GMT</pubDate>
            <atom:updated>2025-02-18T17:04:54.865Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*stDgHyMjm0e6oMwpw7MKIg.png" /><figcaption><em>Figure. 1 Voice of customer</em></figcaption></figure><p>Improve your service by understanding open-ended survey questions. Use BERTopic, OpenAI models, and Databricks to find patterns and learn what customers need. Also, automate the assignment of topics using Databricks Mosaic AI vector search. Turn surveys into plans for better service.</p><p>Imagine you are selling a product or service. You send out surveys and ask your customers to share their thoughts and problems in their own words. These open-ended questions can give you very valuable information. They can help you see not just what problems exist, but also why they are happening.<br> <br>Now, picture an unsupervised topic modeling technique as a pattern identifying robot. This robot can go through all these answers and help you understand your customers’ needs and expectations better.<br> <br>While the robot is busy sorting out the puzzle, your team has more time. They can use this time to understand the picture the robot is creating. This can help them make better decisions and improve your services.</p><p>This article explains how to automatically analyze review texts. It includes steps like preparing the text, which includes removing offensive language, dividing the text by meaning, cleaning the text, and figuring out the best number of topics using BERTopic and Databricks MLflow. This is done through assessing Topic Coherence. It also talks about using OpenAI LLMs such as gpt-4 to represent the topics and using Mosaic AI vector search on Databricks to automatically assign topics to new feedback.</p><p>Enjoy the results!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ip4tpZty9kYnoK7YFx8J9w.png" /><figcaption><em>Figure 2. Customer reviews clustering workflow</em></figcaption></figure><h3>1. Semantic Chunk Generation with OpenAI LLM model</h3><p>The first step in our approach involves using a state-of-the-art Language Learning Model (LLM) to generate semantic chunks from text reviews. Unlike traditional tokenization, which typically breaks down text into words or simple phrases, semantic chunking divides text into larger, more meaningful segments based on semantic content. For example, the sentence “The company has excellent tech leads, and the organization is building a strong technical path of growth” would be chunked into [“The company has excellent tech leads “, “ the organization is building a strong technical path of growth “].<br> <br>The LangChain framework for LLM has recently added a <a href="https://python.langchain.com/docs/how_to/semantic-chunker/">Semantic Chunker</a> as an option for splitting documents based on semantic similarity using a model such as “Text-embedding-ada-002”. This method divides the data while keeping its full meaning, making sure each section represents a whole idea or concept. It not only maintains the quality of the information but also improves the topic modeling process.</p><p>Additionally, it allows for multi-labeling within texts where two or more distinct topics can be identified, providing a richer analysis of the content. We have configured a Databricks serving endpoint to invoke the OpenAI <a href="https://openai.com/index/new-and-improved-embedding-model/">text-ada-embedding-002</a>. This setup allows us to use Databricks as a gateway (Mosaic AI Gateway), creating a single endpoint for calling the model. This makes it easier to manage and allows us to set a rate limit, controlling the number of calls that can be made per minute. In the code snippet below, the text is then divided into documents using the text_splitter method.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/517/1*iOqoWlFHWrgotDB4WECruQ.png" /><figcaption><em>Figure 3. Utilizing Databricks Serving Endpoint for Semantic Text Chunking</em></figcaption></figure><p>An additional step of chunking based on semantics is required from GPT-4 only for reviews longer than 300 characters. However, this could potentially be applied to all reviews, regardless of their length. With the following prompt, we ask gpt-4 to return the text split based on its semantic meaning.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/798/1*tTKE3fG9FDNlTIgRM5ji0w.png" /><figcaption><em>Figure 4. Semantic Chunking with gpt-4 using customized prompt.</em></figcaption></figure><p>In this implementation, we utilized Databricks pandas UDFs to distribute computation for data preparation across multiple workers. After preparing the semantic chunks, we need to clean them up. We removed emojis, smiley faces, and personal information such as tax numbers, emails, numbers, and highly toxic chunks. We also took out punctuation and stop words and we transformed everything to lowercase. Once cleaned, the texts are used for training the BERTopic model. BERTopic is a system for finding topics that uses language models to understand text like Sentence-BERT (Bidirectional Encoder Representations from Transformers) or GPT (Generative Pre-trained Transformer) of OpenAI.</p><h3>2. Training BERTopic with Semantic Chunks</h3><p>In its default version, BERTopic utilizes embeddings derived from SBERT, combined with c-TF-IDF, to cluster texts into thematically coherent groups. The objective of this study is to use a different language model, such as text-embedding-ada-002. Here’s a brief description of BERTopic components:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/1*w0XT3wJgZ3pf1s_4zi0BkQ.png" /><figcaption><em>Figure 5. BERTopic Steps [Bertopic Documentation]</em></figcaption></figure><h4><strong>Text Embedding with OpenAI transformers</strong>:</h4><p>BERTopic begins by transforming text documents into numerical representations, or embeddings, utilizing the text-embedding-ada-002 model. OpenAI text-embedding-ada-002 is one of the most advanced models for generating text-embeddings or dense representations of text that capture their semantic meaning. Text embeddings is an essential step for grouping documents or text based on their content.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/619/1*Q6CtO7pI_OSYGYg-yvOATw.png" /><figcaption><em>Figure 6. BERTopic uses OpenAI’s Transformers as an Embedding Model</em></figcaption></figure><p>The following paragraph (Transformers Architecture), covering transformers architecture to feed forward, is a side note to understand the transformer architecture that BERTopic uses to generate embeddings.</p><p>Feel free to skip if you are already familiar with the topic.</p><h4><strong>Transformers Architecture</strong></h4><p>The architecture used in BERTopic to generate embeddings is primarily based on the “Encoder” component of the Transformer. It consists of a series of stacked encoder blocks, each of which contains two main subcomponents: <strong>Self-Attention Layer </strong>and<strong> Feed-Forward Neural Network</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/355/1*Usv8rAajzn-3I_rDUWjBqg.png" /><figcaption><em>Figure 7. Transformers Architecture for embedding generation [Attention Is All You Need]</em></figcaption></figure><p>Here’s a brief description of its architecture: input sequences are used to create queries, keys, and values. The query is what I am looking for in other tokens, the key is the importance of the token in answering the query, and the value is the informational content of the token.</p><p><strong>Input Embedding</strong>: initially, each piece of input text — typically sentences or documents — is split into pieces: chars, words, or tokens. Tokenization breaks text into smaller units called tokens. A subword-level tokens example:</p><p>“The child plays.”</p><p>[“The”, “child”, “plays”]</p><p>Tokens are indices into a vocabulary (32K)</p><p>[“The”, “child”, “plays”]</p><p>[4 56 676]</p><p>Each vocab entry is mapped to a learned d-dimensional vector (high-dimensional vectors 1536 for text-embedding-ada-002). For simplicity, let’s assume it is 4.</p><p>[4 56 676] -&gt; [[0.1, 0.3, 0.5, 0.7], [0.2, 0.6, 0.8, 0.5], [0.4, 0.9, 0.5, 0.2]]</p><p>Positional encoding is to keep track of the position of different token:</p><p>(“The bird follows the butterfly” is different from “The butterfly follows the bird”)</p><p>[The] + 10 + [child] + 20 + [plays] + 30</p><p><strong>Multi-Headed Self-Attention</strong>: For each token, we calculate its attention score in relation to the other tokens. The values are weighted using the attention weights, producing a new representation of each token.</p><p>“<strong>The</strong>”:</p><p>Query (Q) = Wq ⋅ [0.1,0.3,0.5,0.7]→ [0.2,0.4,0.6,0.8]</p><p>Key (K) = Wk​⋅ [0.1,0.3,0.5,0.7] → [0.1,0.2,0.3,0.4]</p><p>Value (V) = Wv​⋅ [0.1,0.3,0.5,0.7] → <strong>[0.3,0.5,0.7,0.9]</strong></p><p>“<strong>child</strong>”:</p><p>Query (Q) = [0.4,0.6,0.8,1.0]</p><p>Key (K) = [0.3,0.6,0.9,1.2]</p><p>Value (V) = <strong>[0.6,0.7,0.8,0.9]</strong></p><p>“<strong>plays</strong>”:</p><p>Query (Q) = [0.5,0.7,0.9,1.1]</p><p>Key (K) = [0.2,0.4,0.6,0.8]</p><p>Value (V) = <strong>[0.7,0.8,0.9,1.0]</strong></p><p><strong>Computing Attention Scores: </strong>the attention mechanism within the encoder computes scores to determine the influence of each part of the input on the final embedding. These scores are calculated using a function, such as the dot product, between the query and all keys generated from the input sequence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1002/1*FwVruzyIfCe4okLko6a9UQ.png" /><figcaption>Figure 8. Dot product Calculation for Attention Scores</figcaption></figure><p><strong>SoftMax to Weigh Importance</strong>: the attention scores are then normalized using a SoftMax function, which turns them into a distribution of weights summing to one. This step emphasizes certain parts of the input while diminishing others, based on their relevance to the current output token being generated.</p><p>«The»: softmax(0.7)=0.2</p><p>«child»: softmax(1.74)=0.6</p><p>«plays»: softmax(1.2)=0.2</p><p><strong>Creating the Context Vector</strong>: The normalized attention scores are used to create a weighted sum of the values within the encoder itself. This results in what can be considered a context vector for each token. However, instead of preparing to generate the next word in a sequence, this context vector enhances the representation of the entire sentence. Each token’s updated embedding, influenced by the context vector, contributes to the final sentence embedding. This embedding effectively captures the semantic essence of the entire input, optimizing the sentence for tasks that involve measuring textual similarity and relevance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/577/1*vhYR-v7NoT52ATOmOPXQJA.png" /><figcaption>Figure 9. Computing the Context Vector Using Weighted Values</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/426/1*J4rJGor0Ge9Yz4k9R3_ypQ.png" /><figcaption><em>Figure 10. Self-Attention [Lecture: Practical Applications of Large Language Models with PyTorch and Hugging Face]</em></figcaption></figure><p><strong>Feed Forward</strong>: feed-forward networks are a Multilayer perceptron (MLP) applied to each token. It consists of two linear transformations with a non-linearity (ReLU) in between. Before reaching the feed-forward layers, each token’s representation is adjusted through self-attention mechanisms, each token ponders for itself about what it has observed previously. MoE (Mixture of Experts) is applied. MoE, is a specialized subnetwork designed for handling complex tasks by dividing them into simpler sub-tasks.</p><p><strong>Add and Norm</strong>: It is only a technical layer that improves the stability and effectiveness, allowing outputs from earlier layers to be added directly to deeper layers.</p><h4><strong>Dimensionality Reduction with UMAP</strong></h4><p>The embeddings generated by language representation models are typically high-dimensional. Managing high-dimensional data can be complex and computationally intensive. To address this, the high-dimensional embeddings are reduced to a lower-dimensional space using Uniform Manifold Approximation and Projection (UMAP). UMAP aids in preserving both the local and global structure of the data during the dimensionality reduction process.<br>Calculating the embeddings <strong>prior</strong> to the fit_transform operation could enhance the efficiency of the model by reducing the time required for topic detection: precomputing the embeddings avoids the need to recalculate them every time fit_transform is executed, thereby reducing computational load. Moreover, utilizing UMAP for dimensionality reduction can potentially improve the performance and interpretability of the resulting topic model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/842/1*GrsG1I-V_J8B0KsxJM99kA.png" /><figcaption><em>Figure 11. Precomputing embeddings to use in BERTopic</em></figcaption></figure><h4><strong>Clustering with HDBSCAN</strong></h4><p>The reduced embeddings are clustered using Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN). HDBSCAN identifies dense regions in reduced space and assigns them to clusters. In this implementation, we use the following parameters:</p><ul><li>n_gram_range (the n-gram range for the CountVectorizer)</li><li>min_cluster_size (minimum size of a cluster)</li><li>n_components (dimension of embeddings)</li><li>n_neighbors (number of neighboring points used in local approximations of manifold structure)</li><li>min_samples (that controls the number of outliers)</li><li>min_df (to filter out words that have a minimum frequency)</li></ul><p>We first tune the above parameters to maximize NPMI (Normalized Pointwise Mutual Information), a measure of <a href="https://github.com/MIND-Lab/OCTIS">Topic coherence</a> that measures how semantically similar the top words (typically top 10 or top 20) in a topic are. A higher topic coherence score indicates that the top words in a topic are related to each other, suggesting that the topic is more interpretable and makes sense to humans, and topic diversity that measures how different the top words of each topic are from each other. Higher topic diversity means that the topics are more distinct from each other, which could make it easier for humans to differentiate between them.</p><p>To determine the optimal number of topics, we run the model using various topic counts and calculate the topic coherence and topic diversity for each configuration. The numbers of topics that yield the highest topic coherence are considered the optimal numbers for the two-level of topics.</p><p>We used BERTopic’s reduction feature to generate a parent and child level of clusters, thus achieving a two-level hierarchical clustering for easy understanding and visualization of the data clusters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/530/1*rNepyw1DPwK_iXS1YLEsSw.png" /><figcaption><em>Figure 12. Reduction of topics</em></figcaption></figure><h4><strong>Tokenization and Count Vectorization</strong></h4><p>In this step, the original documents are broken down into individual words or tokens, a process known as tokenization. After tokenization, the CountVectorizer is used to convert these tokens into a document-term matrix. In this matrix, each row corresponds to a document and each column represents a unique word from all the documents. The values in the matrix represent the frequency of each word in each document.<br>The parameter <em>min_df</em> is set to filter out words that have a minimum frequency. This means that only the words that appear at least min_df times across all documents will be considered. This helps to remove rare words which may not contribute much to understanding the main topics. The parameter ngram_range is set to (1,2). This means that the CountVectorizer will consider individual words (1-grams) as well as pairs of consecutive words (2-grams). This can help capture more meaning than considering only individual words.</p><figure><img alt="CountVectorizer" src="https://cdn-images-1.medium.com/max/714/1*YkDiQHEWEpMAZDxMqHV06Q.png" /><figcaption><em>Figure 13. CountVectorizer</em></figcaption></figure><h4><strong>Topic Representation</strong></h4><p>A basic part of BERTopic is how it uses the Bag-of-Words model, which is improved by weighting with c-TF-IDF (Class-based Term Frequency-Inverse Document Frequency). Normally, TF-IDF measures the significance of words across different documents. By considering all documents within a single cluster as one collective document and calculating TF-IDF accordingly, we end up with significance scores for words specific to that cluster. Extracting the most significant words from each cluster provides clear descriptions of the topics. The <strong>c-TF-IDF</strong> score can be defined:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WlZi6RMYj-JgdVz7SfrH-g.png" /><figcaption><em>Figure 14. Class-based TF-IDF [BERTopic documentation]</em></figcaption></figure><h4><strong>Create easy-to-understand labels using OpenAI’s text generation models</strong></h4><p>This method uses Bag-of-Words and c-TF-IDF techniques to quickly produce keywords for each topic, separate from the clustering process. We then provide these keywords and a set number of key documents for each cluster to the text generation model (gpt-4). The model’s job is to generate text that closely matches the topic.</p><figure><img alt="Custom Prompt OpenAI" src="https://cdn-images-1.medium.com/max/787/1*yft7ogqbgniImF4KhgnL_Q.png" /><figcaption><em>Figure 15. Custom prompt to create understandable labels for topics.</em></figcaption></figure><h3><strong>3. Filtering documents with high cluster membership probability: outlier Re-evaluation for Detecting Emerging Themes</strong></h3><p>As documented, BERTopic assigns a topic to each document in the dataset along with a probability score. This score represents the probability of the model that a particular document belongs to the assigned topic. After this assignment, a filter could be applied to the dataset such that only documents with a probability score of 0.8 or higher are retained. This means that only documents for which the model has high probability (80% or more) in their topic assignment are kept in the dataset. The remaining documents with a probability score less than 0.8 could be considered outliers. This means that the model is less confident about the topic assignment for these documents, and they are thus treated as not fitting well into any of the identified topics. Indeed, the documents that are marked as outliers will subsequently be used for periodic clustering to identify emerging clusters. This means these outlier documents, which initially didn’t fit well into the identified topics, will be periodically re-evaluated. This re-evaluation or periodic clustering of outlier documents will also consider new documents that are added to the dataset. This means that both the existing outlier documents and the incoming new documents will be analyzed together. As an additional step to BERTopic, we will explore in the next article how to use MLFlow’s evaluation feature to label outliers and identify new topics using a different LLM as a judge.</p><h3><strong>4. Create Stable Topics with Pre-calculated Embeddings and Mosaic AI Hybrid Vector Search:</strong></h3><p>Turning chunks of text into embeddings is important to reinforce the topics identified by BERTopic and OpenAI gpt-4. These embeddings are helpful for a Databricks Hybrid Vector Search where we find documents based on their meaning and matching keywords. They also help us see how similar two chunks of text are. A Delta table is created to store pre-calculated embeddings. A vector index is specifically created on the embedding’s column, which gets synchronized each time the data changes. This synchronization occurs after new customer reviews are added, chunked, cleaned, and classified under a topic different from the outlier topic. This data is copied into the Mosaic AI Vector Search. This vector search is a vector database built within the Databricks Data Intelligence Platform. Newly collected, cleaned, and chunked feedback does not initially belong to any cluster. Through a hybrid vector search, combining vector-based embedding search with traditional keyword-based search techniques, we retrieve text that is not only exact in keyword matches but also conceptually similar. A daily Databricks Workflow is scheduled to assign the same parent and child labels or clusters if a certain similarity threshold (0.9) is met. As we start with an unlabeled dataset and use an unsupervised approach, it is possible that not all feedback is clustered correctly. Therefore, we ask users to provide input only on feedback they believe has been misclassified. Based on this input, we update the Mosaic AI Vector Search index and enhance the inference of new feedback.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DGzYHh75MvymLOVvUoW6FA.png" /><figcaption><em>Figure 14. Mosaic AI Vector Search [Databricks]</em></figcaption></figure><h3>5. Improving Data Analysis with LLMs and MLFlow: Topic Naming, Sentiment Analysis, and Model Deployment</h3><p>By using advanced Language Model (LLM) systems like GPT-4 and employing prompt engineering techniques, we significantly improved the labels for topic naming in our data analysis. Initially, the output table included customer review text, topic, and a brief description. Now, it features additional elements like sentiment analysis, providing a deeper and more detailed understanding of the data.<br> <br>Additionally, we set up a system to create regular summaries of reviews for each topic, whether weekly or monthly. These summaries offer a quick overview of the discussions on each topic during a specific period, helping us easily spot trends, potential issues, or areas of interest.<br> <br>We also adopted MLFlow, an open-source platform designed to manage the entire lifecycle of machine learning projects. This tool helps us effectively monitor and manage our LLM applications, which encompass tasks like topic labeling, sentiment analysis, and summarization. We integrated a PyFunc model in MLFlow as a wrapper for the BERTopic model to centralize and streamline sharing across our processes. MLFlow provides capabilities for tracking experiments, packaging code into reproducible runs, and sharing and deploying models. This platform supports a systematic, consistent approach to evaluating and deploying models.<br>In the next article, we will explore in detail how we used MLFlow to assess the performance of the LLM models we used for sentiment analysis and topic labeling.</p><p><strong>References</strong>:</p><p>1. Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova — <a href="https://arxiv.org/abs/1810.04805"><strong>BERT</strong>: <strong>Pre-training of Deep Bidirectional Transformers for Language Understanding</strong></a></p><p>2. Maarten Grootendorst: <a href="https://maartengr.github.io/BERTopic/index.html"><strong>BERTopic documentation</strong></a></p><p>3. Maarten Grootendorst: <a href="https://arxiv.org/abs/2203.05794"><strong>BERTopic: Neural topic modeling with a class-based TF-IDF procedure</strong></a></p><p>4. <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Vaswani,+A">Ashish Vaswani</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Shazeer,+N">Noam Shazeer</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Parmar,+N">Niki Parmar</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Uszkoreit,+J">Jakob Uszkoreit</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Jones,+L">Llion Jones</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Gomez,+A+N">Aidan N. Gomez</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Kaiser,+L">Lukasz Kaiser</a>, <a href="https://arxiv.org/search/cs?searchtype=author&amp;query=Polosukhin,+I">Illia Polosukhin</a>: <a href="https://arxiv.org/abs/1706.03762"><strong>Attention Is All You Need</strong></a></p><p>5. Kevin Roitero — Lecture:<strong> Practical Applications of Large Language Models with PyTorch and Hugging Face</strong></p><p>6. Silvia Terragni, Elisabetta Fersini, Bruino Giovanni Galuzzi, Pietro Tropeano, Antonio Candelieri:<strong> </strong><a href="https://github.com/MIND-Lab/OCTIS"><strong>OCTIS — Comparing and Optimizing Topic</strong> models Is Simple</a></p><p>7. Databricks:<strong> </strong><a href="https://docs.databricks.com/en/generative-ai/vector-search.html"><strong>How does Vector Search work?</strong></a></p><p>8. Databricks<strong>: </strong><a href="https://docs.databricks.com/en/generative-ai/create-query-vector-search.html"><strong>Create and configure an endpoint to serve the embedding model</strong></a></p><p>9. LangChain: <a href="https://python.langchain.com/docs/how_to/semantic-chunker/"><strong>How to split text based on semantic similarity</strong></a></p><p>10. OpenAI: <a href="https://openai.com/index/new-and-improved-embedding-model/"><strong>New-and-improved-embedding-model</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aff3d4679011" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/transforming-voice-of-customer-into-actionable-insights-with-bertopic-and-openai-on-databricks-aff3d4679011">Transforming Voice of Customer into Actionable Insights with BERTopic and OpenAI on Databricks</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Boundaries and Opportunities: Agile Coach and Corporate Coach Meet]]></title>
            <link>https://medium.com/eni-digitalks/boundaries-and-opportunities-agile-coach-and-corporate-coach-meet-c2ff66f2b774?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/c2ff66f2b774</guid>
            <category><![CDATA[coaching]]></category>
            <category><![CDATA[agile-coach]]></category>
            <category><![CDATA[collaboration]]></category>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[icf-coaching]]></category>
            <dc:creator><![CDATA[massimo cantu']]></dc:creator>
            <pubDate>Fri, 31 Jan 2025 17:01:51 GMT</pubDate>
            <atom:updated>2025-01-31T17:01:51.258Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*OgFzFXS2xFWbv9sMLlhntg.png" /><figcaption>Figure 1: lovelyday12 — stock.adobe.com — AdobeStock_408069086</figcaption></figure><p>This article has been prepared with the valuable input of <a href="https://medium.com/@clod.seb">Claudia Sebastiani</a></p><p><em>What do corporate coaches and Agile coaches have in common? What skills need to be developed? When did these roles emerge? These topics will be addressed and explored by those who hold these positions.</em></p><p>If you are considering <strong>a path to becoming a coach</strong>, it is likely that you are interested in understanding how this role can make a difference both personally and within organizations.</p><p>For instance, if you are wondering which <strong>contexts an Agile Coach can be most effective in</strong>, and where a Corporate Coach might offer greater contribution, here are some reflections to enrich this discussion.</p><p>Although these two roles have distinct goals and approaches, they both play a crucial role in modern organizations, especially in an era where change and innovation are the order of the day.</p><p>So, <strong>what is their impact on businesses</strong>? The Agile Coach stands out for supporting teams and projects, promoting the adoption of agile methodologies to improve collaboration and productivity. The Corporate Coach helps leaders and individuals develop the competencies needed to achieve shared goals.</p><p>Understanding which role is best suited to different business needs is a key step in <strong>fully harnessing the potential of coaching</strong>.</p><p>The idea for this article originated from a conversation by the coffee machine, where we shared our past and present experiences, and one word in particular caught our attention: the word was COACHING.</p><p><strong>WHO IS AND WHAT DOES A CORPORATE COACH DO?</strong></p><p>What is coaching? <strong>ICF</strong> (<a href="https://www.coachingfederation.it/">International Coaching Federation</a>) defines coaching as a <strong>partnership with clients that, through a creative process, stimulates reflection and inspires them to maximize their personal and professional potential.</strong></p><p>Each client is seen as a creative individual full of resources. Coaching supports the exploration of new perspectives and personal opportunities, fostering the development of thinking and decision-making abilities. It also aims to enhance effectiveness in interpersonal relationships and strengthen confidence in living and interpreting one’s roles with awareness both in work and life.</p><p>What does a Coach do? The main responsibilities of a coach include:</p><ul><li>Supporting the identification, understanding, and <strong>alignment of the goals</strong> the client aims to pursue.</li><li>Guiding the client through a <strong>personal discovery process</strong> related to these goals.</li><li><strong>Stimulating the emergence of solutions</strong> and strategies directly from the client.</li><li>Ensuring <strong>full autonomy and responsibility</strong> for the client’s growth process.</li></ul><p>In coaching, we look at “where the client is today,” the current situation, and, by mutual agreement, define what the client is willing to do to reach “the goal they would like to be at tomorrow,” establishing meaningful result indicators.</p><p>These fundamental aspects, integrated with emerging themes, today represent <strong>the pillars of coaching practice</strong> and will serve as standards for the future of the profession.</p><p>Below are the 8 core competencies that the ICF coach constantly develops and practices:</p><ol><li><strong>Demonstrates Ethical Practice</strong><br> Understands and consistently applies coaching ethics and standards; demonstrates personal integrity and honesty in interactions with clients, sponsors, and stakeholders.</li><li><strong>Embodies a Coaching Mindset</strong><br> Develops and maintains an open, curious, flexible, and client-centered mindset.</li><li><strong>Establishes and Maintains Agreements</strong><br> Collaborates with the client and stakeholders to create clear agreements on the relationship, process, plans, and coaching goals. Establishes agreements for the entire coaching journey as well as for each individual session.</li><li><strong>Cultivates Trust and Safety</strong><br> Works with the client to create a safe and supportive environment that allows the client to share freely. Maintains a relationship of mutual respect and trust.</li><li><strong>Maintains Presence</strong><br> Is fully aware and present with the client, employing an open, flexible, centered, and self-confident style.</li><li><strong>Listens Actively</strong><br> Focuses on what the client says and doesn’t say to fully understand what is being communicated in the context of the client’s systems and to support the client’s self-expression.</li><li><strong>Evokes Awareness</strong><br> Facilitates the client’s insights and learning using tools and techniques such as powerful questions, silence, metaphors, or analogies.</li><li><strong>Facilitates Client Growth</strong><br> Collaborates with the client to transform learning and insights into actions. Promotes the client’s autonomy in the coaching process.</li></ol><p><strong>WHO IS AND WHAT DOES AN AGILE COACH DO?</strong></p><p>The Agile Coaching Institute <strong>(ACI)</strong> was founded in 2010, officially establishing the role of the Agile Coach. This event is also considered by Charlie Rudd, in his famous article The Third Wave of Agile, as one of the key milestones that marked the end of the first wave of Agile.</p><p>In the <a href="https://agilecoachcompetencyframework.com/"><strong>Agile Coach Competency Framework</strong></a><strong> </strong>developed by <a href="https://lyssaadkins.com/">Lyssa Adkins</a> and <a href="https://www.linkedin.com/in/spayd/">Michael Spayd</a>, as members of the Agile Coaching Institute, the 8 key competencies of the Agile Coach are represented:</p><ol><li>Agile-Lean Practitioner</li><li>Teaching</li><li>Mentoring</li><li>Professional Coaching</li><li>Facilitating</li><li>Technical Mastery</li><li>Business Mastery</li><li>Transformation Mastery</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1015/1*hL7souRKK3je8IwXMt-9BQ.png" /></figure><p>The first point, which is also highlighted at the top of the framework, <strong>Agile-Lean Practitioner</strong>, reminds us that the Agile Coach, above all, embodies the Agile mindset and is the first to apply its practices.</p><p>On the left side of the framework, we find two competencies: <strong>the ability to teach</strong>, by delivering courses on Agile practices, and <strong>the ability to mentor</strong>, by sharing one’s experience and supporting the person or team in a “training on the job” mode.</p><p>At the bottom, three specialist competencies are described: <strong>Technical, Business, and Transformation.</strong></p><p>The right side contains two other competencies: <strong>the ability to act as a Coach</strong>, leaving control to the coachee, and <strong>the ability to facilitate</strong>, remaining neutral on the content and allowing the team to define goals and processes.</p><p>The fact that Teaching and Mentoring are represented in one area, while Coaching and Facilitating are represented in the diametrically opposite area, is a key point.</p><p>Borrowing from the metaphor of “thinking hats” introduced by Edward de Bono in his book Six Thinking Hats, we can say that the Agile Coach can wear four different hats.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JbfPPh96oqmR3dn5ZAiqig.jpeg" /></figure><p>When the coach wears the <strong>teacher or mentor hat</strong>, they act as a guide in adopting the methodology, working alongside the individual or team in a “training on the job” mode, sharing their experience, and providing solutions to problems that arise.</p><p>On the other hand, when wearing the <strong>coach or facilitator hat</strong>, the coach asks questions to develop awareness and help the team generate solutions on their own, without ever providing answers. Their focus is on the dynamics of the team, aiming to <strong>develop thinking and decision-making abilities</strong> and collaborate more effectively. They apply Agile practices <strong>to build trust and create a safe environment</strong>, without intervening in what the team is creating.<br> So, these are two different orientations: one towards content, the other towards processes.</p><p>In the image below, we can see how content expertise, which at first is entirely in the hands of the coach wearing the teacher/mentor hat<strong>, gradually shifts to the coachee/team</strong> as the process progresses. The coach then assumes the role of facilitator and awareness developer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mOExk1jgHJWDvwIYqQOg6g.jpeg" /></figure><p>Fabio Lisca, founder of <a href="https://agile-school.com/chi-siamo">Agile School Italia</a>, describes this role as one that enhances people’s skills to improve results through the Agile mindset and approaches, while Agile Coaching is seen as an approach to organizational improvement.</p><p>We can also say that the Agile Coach is the one who, without pushing or pulling, acts as a change agent in transformation processes.</p><p><strong>The Path to Becoming an Agile Coach</strong></p><p>To become an Agile Coach, it is necessary to cultivate and develop a strong Agile mindset, practical experience with the four hats, and the development of specific specialist skills. The starting point can be the role of Scrum Master, Project Manager (PM), or even Corporate Coach.</p><p><strong>WHAT WE HAVE LEARNED FROM EACH OTHER</strong></p><p>We can say that both the corporate coach and the Agile coach focus on strategic aspects and act as change agents in transformations.</p><p>Both approaches <strong>start with observing</strong> the current situation, whether it concerns an individual, a team, or an organization, and <strong>establish a mutual agreement</strong> where the “client” sets the “goal they want to achieve.”</p><p>The proposed journey is designed to support the exploration of new perspectives and opportunities, and to foster the development of thinking and decision-making abilities in the individual or team.</p><p>As some tips and tricks, we can say that even when discussing organizational improvement, <strong>it would be a best practice to formalize the agreement</strong> — contract — pact between the coach and the team in order to lay the foundation for the relationship, set the objective of the journey, define the collaboration mode, and come to mutual agreements.</p><p>When <strong>choosing the right “hat,”</strong> it is important to avoid making too broad a leap or keeping the “lowest” hat for too long compared to the maturity level of the organization/team/individual, so as to build an effective journey.</p><p>We conclude by saying that, in discussing the writing of this article, we stumbled upon many other interesting topics. Stay tuned!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c2ff66f2b774" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/boundaries-and-opportunities-agile-coach-and-corporate-coach-meet-c2ff66f2b774">Boundaries and Opportunities: Agile Coach and Corporate Coach Meet</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Master MLOps with Databricks: a comprehensive guide]]></title>
            <link>https://medium.com/eni-digitalks/master-mlops-with-databricks-a-comprehensive-guide-94b2b8140594?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/94b2b8140594</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[databricks]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[Mazzullo Alessandro]]></dc:creator>
            <pubDate>Fri, 24 Jan 2025 14:56:21 GMT</pubDate>
            <atom:updated>2025-01-24T14:56:21.450Z</atom:updated>
            <content:encoded><![CDATA[<p>An in-depth look at CI/CD pipelines, with a focus on Databricks tools and testing strategies to ensure code quality, security, and seamless deployment</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*dN9VImEY38SthB50AnYWsA.png" /><figcaption>Figure 1 — Source Adobe Stock</figcaption></figure><p>This article was written with <a href="https://medium.com/u/af49d46be723">Lavinia Guadagnolo</a></p><h3>Introduction</h3><p>In the implementation of a company’s digital transformation, there inevitably comes a time when the area dealing with data must evolve. When the data team grows over time, it leads to a large number of operational use cases and a larger team of data scientists and engineers engaged in their development. This expansion presents several challenges.</p><p>One pressing issue is the lack of a uniform approach to <strong>collaborative</strong> <strong>project</strong> <strong>work</strong>. Data scientists and engineers often come from diverse backgrounds, leading to inconsistent methodologies and inefficiencies. Another critical need is to improve <strong>time-to-market</strong> by streamlining workflows, ensuring that projects move seamlessly from experimentation to production.</p><p>As the number of projects increases, the complexity of codebases often grows, making it essential to simplify the <strong>integration</strong> of new models. Teams must focus on business-driven use cases rather than administrative and management tasks.</p><p>The solution to these challenges lies in MLOps (Machine Learning Operations), which is a set of practices and tools aimed at streamlining machine learning model development and deployment. It ensures effectiveness in model maintenance and reliability in scaling.</p><p>In this article we will explore a possible practical way of implementing it. Although the illustrated structure is based on Azure DevOps and Databricks, the principles we discuss can offer valuable insights regardless of your specific infrastructure and tools. The aim is to offer guidelines for data scientists, data engineers, and machine learning engineers who are facing similar challenges and looking to embark on their journey into the world of MLOps.</p><p>The framework illustrated in this article has also been presented and discussed at the Data &amp; AI Summit 2024 in San Francisco, the recording can be found <a href="https://www.databricks.com/dataaisummit/session/efficient-mlops-developing-and-deploying-ml-models-databricks-repeated">here</a>.</p><h3>Tools</h3><p>Any MLOps framework can be considered as a combination of tools and practices used to streamline and automate the deployment, monitoring, and management of machine learning models in production environments.</p><p>We will first focus on the Databricks tools we considered crucial. We won’t deep dive into the tools’ description, as you can find specific documentation on their <a href="https://docs.databricks.com/">website</a>, but we’ll mainly focus on how we leveraged these tools in building our framework.</p><h4>Unity catalog</h4><p><a href="https://www.unitycatalog.io/">Unity Catalog</a> is an all-encompassing governance solution designed to improve the <strong>security</strong>, <strong>administration</strong> and <strong>discovery</strong> of data and AI assets in Databricks environments. By establishing a centralized place for managing metadata and assets — including governance protocols, access control, auditing, and data lineage tools — Unity Catalog strengthens security mechanisms within Databricks workspaces.</p><p>Multiple separate instances of Unity Catalog can be deployed to support several target environments such as sandbox, staging and production environments. This architecture guarantees access to production data in a controlled and centrally managed manner, facilitating development and testing while respecting strict security measures and cost efficiency.</p><p>In this way, production data can be leveraged for meaningful testing early in the life and development of models, enforcing segregation and security standards.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/643/1*Ax4FDbOAPjpUNPshRBqd7Q.png" /><figcaption>Figure 2 — Unity Catalog description</figcaption></figure><p>In Figure 3 the icons represent the professional figures located in the environments where they can interact with the assets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/199/1*C06nORCjoLycDuGlIuLXJQ.png" /><figcaption>Figure 3 — Roles legend</figcaption></figure><p>The Unity Catalog is a container for all the assets and artifacts that are involved in the Data Science lifecycle, ranging from models’ versioning and serving to data outputs. The catalogue is flexible in the way it can be organized (Figure 4); for instance, there could be catalogues for business teams, with specialized schemas aligned to the particular data products managed by the teams, models, and tables.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/964/1*XfTWVk5q0DCWHhOsLCxweg.png" /><figcaption>Figure 4 — The hierarchy of database objects in any Unity Catalog metastore</figcaption></figure><p>Thanks to Unity Catalogs, both data and models are orchestrated and governed within a single governance framework. This framework fosters a <strong>data-centric paradigm</strong> that is fundamental to the advancement of machine learning initiatives.</p><h4>Databricks Asset Bundles</h4><p>Managing data products on Databricks can be quite challenging. For each product there are a variety of components. The needed components depend on the specific goals of the data product, and, as the tasks evolve those components must be adjusted accordingly. Additionally, the way the code is run varies according to the environment, requiring careful management of environment-specific configurations.</p><p>Databricks Asset Bundles (DAB) simplify this complexity and make managing these elements more efficiently.</p><p>Let’s see how.</p><p>DAB provide a way to include <strong>metadata</strong> alongside project’s source files. Bundle metadata is defined using YAML files. These YAML files specify the artifacts, resources (including Unity Catalog resources), and configurations of a Databricks project and can be created manually or using a template.</p><p>Once everything is set up, project’s source files and metadata are packaged together and deployed as a single bundle to the target environment, streamlining the deployment process.</p><p>DABs main advantage is to ease the management of complex projects during active development, providing CI/CD capabilities in the software development workflow with a single concise declarative YAML syntax.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/699/1*oxcWozWFhc8YSzq0p1x_Kg.png" /><figcaption>Figure 5 — Repository structure</figcaption></figure><p>In Figure 5 a potential repository structure is shown, where assets are leveraged.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/446/1*eAMx4bIduiizyZ6hw4T_Fg.png" /><figcaption>Figure 6 — model-workflow-asset.yaml</figcaption></figure><p>In Figure 6 a bundle is defined. It is possible to specify the bundle name as well as the url and target workspaces for each stage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/497/1*VckflQk7-urdjCZI0fgO8Q.png" /><figcaption>Figure 7 — batch-inference-worflow-asset.yaml</figcaption></figure><p>In DAB it is also possible to specify Databricks jobs, the clusters in which tasks should be executed and set dependencies between different tasks, to define workflows (Figure 7).</p><p>DAB support <strong>collaboration</strong> by enabling isolation and customization of tests and release steps on a shared workspace. They are also particularly valuable for <strong>compliance and documentation</strong> reasons as they maintain a versioned history of code and infrastructure. By<strong> automating</strong> tests, deployments and configuration management DAB minimize errors and <strong>reduce</strong> <strong>time to market</strong>.</p><p>To conclude, DAB democratizes the process of getting assets into production, providing a <strong>standardized</strong> way to deploy with configurations, while Unity Catalog provides <strong>centralized</strong> and <strong>optimal</strong> <strong>control</strong> and management of assets such as tables and models across environments.</p><p>With these two tools in mind, we can now deep dive into the core of our framework.</p><h3>CI/CD Framework within MLOps</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/825/1*EumwbMaLcWZMNyRrVNJEag.png" /><figcaption>Figure 8 — High level CI/CD process</figcaption></figure><p>In the world of software development, two acronyms have emerged as the main pillars of optimization: CI (Continuous Integration) and CD (Continuous Delivery/Deployment). When combined with MLOps — a fusion of machine learning, development, and operations — these methodologies improve the efficiency and effectiveness of developing machine learning models.</p><p>CI/CD optimizes the process by which developers integrate, test, and deploy code, ensuring that changes are delivered to end users in a smooth and timely manner. Continuous Integration ensures that code changes are automatically incorporated and evaluated within a shared repository, while Continuous Deployment automates the process by releasing code into production as soon as it passes all tests.</p><p>The integration of CI/CD into MLOps offers significant benefits, including a reduction in mean time to resolution (MTTR), increased test efficiency and the ability to focus on innovation for engineers. This methodology transforms software development into a more agile, efficient, and resilient process, paving the way for a leaner digital ecosystem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/643/1*qIQm8HWgFSm6dOQowrKfyg.png" /><figcaption>Figure 9 — Continuous integration pipeline schema</figcaption></figure><p>In the following paragraphs we will explore how the two pipelines can be constructed, but first we will explain how the developers’ work changes with the introduction of this framework. As shown in Figure 9, when a developer makes a change, after evaluating it in the development environment (sandbox), he creates a pull request (PR) with the proposed changes. Already at the time of creation of the PR, some checks are performed locally, concerning the quality of the code, and when these are successfully completed, the CI is automatically started. During the CI, new tests are performed to check whether the change meets the required standards, and then the artefact is built and versioned.</p><p>Artefacts are packages generated from source code, which can be versioned for better tracking and management. They can come in various formats, such as Docker images, binary files, or compressed archives (e.g., zip files) containing code. In the presented case, the artifacts are zip files that contain the repository, as Python does not require a compilation step, unlike other languages. These zip files represent the packaged version of the application or project that is ready for deployment or further processing. If, as desired by the developer, the CI is executed without errors, the CD will automatically start, which performs bundle validation, deployment and builds on all the environments sequentially.</p><p>Figure 9 illustrates various tools interacting with the CI/CD pipeline. The following paragraphs will explain their roles and objectives within the Continuous Integration process.</p><h4>Deep dive CI</h4><p>Tests performed during CI Continuous Integration (CI) play a critical role in maintaining the integrity and reliability of machine learning workflows. These tests go beyond just checking code quality — they ensure that every change made to the codebase does not negatively impact the existing functionality , which is especially important for complex data products.</p><p>In the following table we will illustrate different types of tests.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/964/1*25A3uZQifs4_agtetDNpvw.png" /><figcaption>Table 1 — Tests</figcaption></figure><p>With multiple people collaborating on the same project, we suggest adopting strict rules for how code should be written and documented. Code quality and security are rigorously evaluated during static code analysis, using various tests and tools to verify compliance and integrity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*G3Hkbb30Mv6NUUuh3mLgAA.png" /><figcaption>Table 2 — Static code analysis</figcaption></figure><p>In our framework, the first 3 tools mentioned in Table 2 are performed at commit phase (via pre-commit tool) and enforced during PRs. The last two ones are comprehensive code checks, which provide a complete report which can be stored and shared.</p><p>To conclude, robust CI testing contributes to the overall reliability and success of your ML operations, reducing the risk of failures and promoting smooth, efficient deployments.</p><h4>Deep dive CD</h4><p>Now that we have a clear understanding of how the CI works, it is time to explore a possible structure of the CD. We will examine a hybrid deployment approach. As Figure 10 shows, we have three environments: P1 (sandboxes), P2 (staging) and P3 (production).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/606/1*vxLHOXh4mNviyBAGDOu0LQ.png" /><figcaption>Figure 10 — CD schema</figcaption></figure><p>When a pull request (PR) is created in the P1 environment, a whole series of unit tests and integration tests are executed as we have said, which ensure that the execution’s flow of the algorithm remains correct. Once the tests have been passed, an email is sent to the maintainer, who should be a senior figure and different from the developer who created the PR.</p><p>The green light on the PR ushers the CD into the staging environment (P2), where the revised code is exclusively deployed. This staging environment, mirroring the production one (P3) in configuration and architecture, becomes the place where the algorithm’s full spectrum of operations — ETL processes, training, and so forth — are performed. The resultant model’s performance is then evaluated against the existing production model, determining its suitability for production purposes. If the model trained for staging exceeds the performance of the current model in P3, and with the approval of the maintainer, the new version is promoted to P3, where it only performs inference operations.</p><h4>Evaluation and validation</h4><p>The process of evaluating and promoting models is critical. It must be verified that model’s performance meets the minimum threshold required to be deemed valuable (validation) and that validated model outperforms previous versions, to ensure that only the best model is promoted to production (P3). Model evaluation can be different from model to model, in terms of data needed to validate it, KPI to be considered and error thresholds. However, the process for deciding which version of a model can be used for production purposes generally the same:</p><ol><li>A new version of a model is trained.</li><li>A threshold is applied to define if the new (challenger) model is valid.</li><li>If the challenger model meets the validation threshold it is compared with the current production (champion) model.</li><li>If the challenger proves to be better with respect to the champion it is then promoted itself as champion and the model is deployed, otherwise the current champion model remains as the chosen one.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/617/1*YVYZohYsRfKbQmyC7wgkvQ.png" /><figcaption>Figure 11 — Evaluation flow chart</figcaption></figure><p>We register models in Unity Catalog via <a href="https://mlflow.org/">MLFlow</a>, and opensource framework that helps in managing the model’s lifecycle, since, as we said at the when describing the tool, catalogs enable us to manage model and data consistently.</p><p>This hybrid deployment framework emerges as a pragmatic solution in scenarios where segregation of environments is rigidly imposed. It allows data scientists to evaluate the performance of models with authentic data within the staging environment, enabling them to make informed decisions on the promotion of models to production, decisions that might otherwise be hampered by authorization constraints in production. Furthermore, this strategy avoids the redundancy of repeating algorithmic segments, a common requirement in code distribution strategies, thus streamlining the overall process.</p><h3>Takeaways</h3><p>We illustrated how we built our framework, tailored to our specific context, guided by three essential principles:</p><ul><li>Standardize</li><li>Scale</li><li>Automate.</li></ul><p>Throughout the process of developing, promoting, and applying our framework we have gained valuable insights.</p><p>Firstly, adopting a MLOps approach requires a <strong>cultural</strong> shift as much as it needs technical skills and implementation. Revolutionizing the way projects are carried out is an <strong>investment</strong>. For successful change, it’s crucial that everyone involved is committed to contributing to this transformation. While the initial effort can be substantial, the benefits of MLOps in the short to medium term far outweigh the costs.</p><p>Secondly, MLOps is a <strong>multidisciplinary</strong> approach involving data science, data engineering, architecture. Building an effective framework necessitates collaboration across different departments to integrate diverse expertise.</p><p>Finally, organizational constraints should not be viewed as barriers to adopting an MLOps practice. Instead, they should be seen as challenges to overcome in the pursuit of a more efficient and effective approach.</p><h3>Sources</h3><p>1. <a href="https://www.databricks.com/product/unity-catalog">https://www.databricks.com/product/unity-catalog</a></p><p>2. <a href="https://docs.databricks.com/en/dev-tools/bundles/index.html">https://docs.databricks.com/en/dev-tools/bundles/index.html</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94b2b8140594" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/master-mlops-with-databricks-a-comprehensive-guide-94b2b8140594">Master MLOps with Databricks: a comprehensive guide</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The role of Focus Groups in shaping Employee Experience and well-being]]></title>
            <link>https://medium.com/eni-digitalks/the-role-of-focus-groups-in-shaping-employee-experience-and-well-being-d67bff57deb0?source=rss----fe373d427779---4</link>
            <guid isPermaLink="false">https://medium.com/p/d67bff57deb0</guid>
            <category><![CDATA[tips-and-tricks]]></category>
            <category><![CDATA[employee-experience]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[research-methods]]></category>
            <category><![CDATA[wellbeing]]></category>
            <dc:creator><![CDATA[Silvia Cardella]]></dc:creator>
            <pubDate>Wed, 15 Jan 2025 09:44:21 GMT</pubDate>
            <atom:updated>2025-01-15T09:44:21.505Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Discover how Focus Groups can help organizations enhance employee satisfaction and workplace well-being.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/756/1*SzQkOyYxUv4CX4n_tAF19A.png" /><figcaption>©Friends Stock — stock.adobe.com</figcaption></figure><h3>Setting the stage for employee well-being</h3><p>In the pursuit of fostering a positive and thriving workplace, businesses are increasingly recognizing the <strong>importance of investing in a culture of employee well-being, cultivating a climate of respect and inclusion</strong> where people “<em>want to show up and not merely need to show up to work</em>” (cfr Jacob Morgan), ultimately fostering talent retention. Indeed, prioritizing the needs of the workforce by ensuring a high level of employee satisfaction throughout their experience is key to increasing productivity, engagement, and overall company performance, contributing to its success.</p><p>A particularly effective approach for boosting workplace well-being involves leveraging <strong>qualitative and quantitative research methods</strong> to gain a deeper understanding of the employee experience, identify their needs and concerns, and then implement targeted interventions to foster a positive impact.</p><p>Our in-house <strong>Design Center of Excellence</strong> team at Eni, an Italian multi-national energy company, frequently collaborates with HR and D&amp;I departments to research employee well-being. By using research methods like Interviews, Discovery Workshops, Shadowing, and Focus Groups, we involve our colleagues in a <strong>collaborative “listening and action” effort</strong>. This approach allows us not only to address their needs but also to uncover opportunities for system improvement, making the organization more attractive and enjoyable for employees.</p><p>This article delves into how Focus Groups serve as a dynamic research method, helping organizations understand employee needs, enhance workplace well-being, and drive meaningful change. Drawing from our experiences and Lessons Learned, we’ll share practical steps, tips, and insights to make this method work for you.</p><h3>1. What is a Focus Group</h3><p>A Focus Group is a <strong>qualitative research method</strong> used to collect insights, understand perceptions, and explore opinions about a specific topic through guided discussions with a group of participants. By fostering <strong>open dialogue</strong>, this method uncovers valuable qualitative data and trends that can inform subsequent decisions and actions.</p><p>In a <strong>Design Thinking framework</strong>, Focus Groups are particularly effective during the <strong>research phase</strong>, as they provide structured opportunities to understand user needs and gather deep insights through open, collaborative discussions. This is especially true in workplace contexts, where Focus Groups can engage employees directly to explore their <strong>collective experiences</strong>, <strong>challenges</strong>, and <strong>expectations</strong> and to stimulate their <strong>creative thinking</strong>.</p><p>In our organization, we’ve reimagined traditional Focus Groups by <strong>incorporating workshop elements and interactive activities</strong>, encouraging creativity and engagement.</p><p>Skilled moderators or <strong>facilitators</strong> guide these sessions impartially, capturing authentic feedback, first-hand perspectives, and observable behaviors in a dynamic yet structured manner.</p><p>The insights gathered are then analyzed and translated into actionable improvements that promote inclusivity, transparency, and overall well-being.</p><blockquote><em>Feedback gathered in Focus Groups is valued and acted upon, contributing to promoting a culture of transparency and inclusivity.</em></blockquote><h3>2. How to organize &amp; run Focus Group sessions: a step-by-step guide</h3><p>Based on our experience, we have developed a structured <strong>framework</strong> to conduct effective Focus Groups:</p><ol><li><strong>Preparation phase</strong>: lays the groundwork for research;</li><li><strong>Execution phase</strong>: facilitates planned activities to meet research goals;</li><li><strong>Follow-up phase</strong>: analyzes findings and translates them into actionable plans.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OU4EyoyKuwVSCtmAM8JT3w.png" /><figcaption>Framework designed by Eni Design Team to visualize the main steps to setup a Focus Group.</figcaption></figure><h3>2.1 Set the basis for the research: what is the scope?</h3><p>As we lay the foundation for our research, here are two key tasks to focus on:</p><ol><li>Defining the research scope;</li><li>Identifying the most appropriate research method, which in this case will be Focus Groups.</li></ol><p>The first step is to clearly define<strong> specific and measurable research objectives </strong>to guide the entire process.</p><p>To articulate the <em>research scope,</em> ask yourself:</p><ul><li>What do we want to achieve?</li><li>What are the key questions we need to answer?</li></ul><p>For example, goals might include understanding employee experiences, identifying challenges, or uncovering unmet needs. These objectives must align with the priorities of organizational stakeholders.</p><h4>PRACTICAL TIP:</h4><p><em>Research is an iterative process, so remain flexible and refine your goals as new insights emerge to keep the research relevant and effective.</em></p><p>Since Focus Groups are ideal for uncovering rich qualitative data, we select this method whenever we want to explore complex topics, generate hypotheses, or validate ideas through participant discussions.</p><h3>2.2 Target the audience: who to involve in the research?</h3><p>Choosing the right participants is critical for collecting valuable insights.</p><p>But how to identify the right target audience?</p><ul><li>The research goal will guide the definition of people we need to involve.</li><li>Define the target audience based on <strong>criteria</strong> like: job roles, demographics, or experiences related to the research topic.</li></ul><p>Depending on your goals, Focus Groups can be:</p><ul><li><strong>homogeneous</strong>, where the goal Is to explore commonalities between participants who share similar characteristics;</li><li><strong>heterogeneous</strong>, where the goal is to explore differences among participants’ perspectives who represent diverse backgrounds.</li></ul><p>Once your audience is defined, the next step is to <strong>draw up the list of potential participants</strong>:</p><ul><li>We share a <strong>call-to-action</strong> through internal communication channels inviting volunteers;</li><li>Among voluntarily nominated colleagues we select participants who meet the criteria, <strong>ensuring diversity of viewpoints</strong> to encourage inclusive dialogue;</li><li>We then divide potential participants into <strong>smaller working groups</strong> (3–10 people) to foster open discussions, and observe the dynamics.</li></ul><h4>PRACTICAL TIP:</h4><p><em>Unless the research context requires so, our advice is to limit managers and top executives from groups involving their subordinates, to mitigate hierarchical barriers and encourage candid contributions.</em></p><h3>2.3 Design focus groups’ approach and logistic: what is the plan?</h3><p>To ensure a smooth session, <strong>design an engaging approach</strong> while <strong>organizing logistics</strong> efficiently.</p><p>Here’s the plan:</p><ul><li>Tailor the session format based on your research goals. We mostly organize Focus Group sessions in a <strong>workshop format</strong> with structured and <strong>interactive activities</strong>, such as brainstorming sessions, scenario analysis, and role-playing, to enhance engagement and foster participants’ creativity.</li><li>When designing the approach, allocate time for <strong>individual reflection</strong>, <strong>group collaboration,</strong> and <strong>plenary discussions </strong>with specific activities.</li><li>Prepare supporting materials, like printed or digital <strong>canvases</strong> and <strong>templates</strong>, in advance to help execute the activities or tasks.</li><li>Depending on the format of the session you’ve determined (whether it will be held <strong>in-person, remotely, or in a hybrid format</strong>) you may need to adjust the <strong>duration</strong> of each activity and, if necessary, the <strong>number of sessions</strong> and <strong>groups</strong> running concurrently.</li><li>Plus, allocate time for introductions, activity explanations, and a wrap-up session to synthesize key findings and draw conclusions.</li></ul><p>Focus on effective preparation; while prioritizing session design, ensure logistical and technical details are also addressed:</p><ul><li>Prepare a <strong>slide deck</strong> to clearly communicate the purpose of the Focus Group session, explain the activities in a detailed <strong>agenda,</strong> and provide participants with any additional instruction.</li><li>If you are planning to organize <em>in-person sessions</em>, <strong>book a comfortable venue</strong>, prepare all necessary materials in advance (e.g., stationery, printed canvases, and physical whiteboards), and curate a collaborative setting.</li><li>If you are planning to organize <em>remote sessions</em>, <strong>set up video conferencing tools</strong> (e.g., Microsoft Teams), organize breakout rooms, and prepare <strong>virtual whiteboards</strong> (e.g., Mural, Miro) in advance.</li><li>You may also want to conduct <strong>multiple sessions concurrently</strong> in separate rooms or spread them out over different days.</li><li>Identify <strong>skilled facilitator(s)</strong> to guide the discussion, explain the activities, and gather contributions during the session development.</li><li>Finally, prepare an <strong>engagement email</strong> to secure participants’ confirmation for the sessions. Once the desired number of participants is confirmed (remember to include some backup names in case invited individuals are unable to attend), <strong>share invitations</strong> and the agenda with participants, providing clear instructions beforehand.</li></ul><h4>PRACTICAL TIPS:</h4><ul><li><em>Divide participants into </em><strong><em>small working groups</em></strong><em> to ensure diverse viewpoints and active involvement. Encourage balanced contributions from all participants by planning moments for individual reflection, followed by group sharing. This approach encourages all participants to share their insights effectively and ensures that no single voice dominates the discussion and influences the entire group.</em></li><li><em>Conduct a </em><strong><em>Dry Run session</em></strong><em> a few days prior to the Focus Group to test tools and timing, and to ensure everything is working correctly.</em></li></ul><h3>2.4 Run the Focus Group session: how to collect data?</h3><p>Moving on to the <em>Execution phase</em>, here is how we typically structure Focus Groups into two key stages:</p><ul><li><strong>Discovery Phase</strong>: participants reflect on their individual experiences and share insights through group discussions. This phase focuses on exploring unmet needs and clustering common themes.</li><li><strong>Definition Phase</strong>: participants collaboratively analyze and prioritize these themes in groups, exploring their interconnections and impact. This helps define actionable opportunities and solutions.</li></ul><p>Let’s explore an example of an agenda:</p><ul><li>Start by providing context on the research goals and session objectives to participants. Clearly communicate the purpose and explain how their participation will contribute to achieving those goals.</li><li>Then, facilitate <strong>needs exploration</strong> by encouraging participants to independently reflect on their experiences and opinions before sharing their insights with the group in open discussions and clustering common insights to consolidate individual needs into shared topics.</li><li>Guide collaborative discussions to converge on <strong>the most significant clusters of needs</strong> and examine their impact and interconnectedness between i.e., individual, organizational, and societal needs, to foster a holistic perspective.</li><li>Finally, encourage participants to <strong>visually synthesize</strong> the insights, concepts, and ideas using visual tools like Personas or journey maps as references for specific needs, opportunities and identified goals.</li><li>Wrap up and facilitate a <strong>final discussion</strong> to convey the key findings, gather feedback, and identify areas of focus for the development of tailored solutions.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5Q0tN1AoRzQyTLzjEtUXSw.png" /><figcaption>Approach example designed by Eni Design Team to visualize how to execute a Focus Group session.</figcaption></figure><h4>PRACTICAL TIPS:</h4><ul><li><em>This approach provides an opportunity to </em><strong><em>observe non-verbal cues</em></strong><em>, </em><strong><em>group dynamics</em></strong><em>, and social interactions among participants, offering richer insights and additional context to the data collected. Ask participants if you can record specific parts of the session through video or pictures as a reference for analysis.</em></li><li><em>Emphasize the </em><strong><em>importance of open and creative thinking</em></strong><em> to foster collaboration and sharing of their experiences, needs, and opinions openly, following the planned activities facilitated by the moderator who guides the discussion and manages the session flow.</em></li></ul><h3>2.5 Analyze collected data: how to deliver key insights?</h3><p>After the session, analyze the collected data to extract <strong>insights</strong> uncover <strong>recurring themes</strong> and <strong>trends</strong>. Pay attention to:</p><ul><li>The <strong>diversity of viewpoints</strong>, to discern trends and areas of consensus;</li><li>The <strong>language</strong> and <strong>tone</strong> used by participants, which can offer deeper context into their beliefs, experiences, and motivations within their employee experience.</li></ul><p>The analysis phase aims to synthesize the raw data into clear and actionable insights that directly address the research objectives. In Eni we summarize the research findings in a <strong>comprehensive report</strong>, which includes:</p><ul><li>An overview of the <strong>research objectives</strong> and the <strong>methodology</strong>;</li><li><strong>key insights</strong> and relevant contextual factors, supported by visual aids (charts, graphs, images);</li><li>a <strong>description of the target</strong> selected to participate in the research, outlining the characteristics identified through defined criteria;</li><li>a <strong>summary of the discussions</strong> that emerged based on the predefined objectives;</li><li>any other pertinent information relevant to the research scope.</li></ul><p>This report represents a valuable resource we share with stakeholders responsible for decision-making, as it provides them with useful insights to plan strategic improvement actions.</p><h4>PRACTICAL TIP:</h4><p><em>Include </em><strong><em>actionable recommendations</em></strong><em> and practical next steps in the report, to ensure all relevant findings are translated into meaningful actions.</em></p><h3>2.6 Take action:</h3><p>In the final phase, key stakeholders who have commissioned the research take the lead in developing action plans.</p><p>As designers, we support this process by integrating insights into <strong>broader organizational strategies</strong>, ensuring that actions are aligned with employee needs and organizational goals.</p><h3>3. Final Takeaways and tips for running effective Focus Groups:</h3><p>To designers and professionals seeking to improve employee experience and workplace well-being, <strong>continuous research</strong> is key. Regularly gathering feedback and insights through methods like Focus Groups, ensures organizations remain attuned to evolving workplace dynamics and proactively address challenges.</p><p>Focus Groups provide a structured method to <strong>understand employee experiences</strong>, <strong>foster belonging</strong> and <strong>empowerment</strong>, and enhance overall satisfaction and engagement. By identifying needs and uncovering pains, organizations can make informed decisions, implement targeted interventions, and create a supportive and favorable work environment that benefits both employees and the company.</p><p><strong>As designers, our role extends beyond facilitation</strong>: we drive innovative, employee-centered solutions by leading engaging sessions and creatively analyzing data. By championing a culture of listening and empathy, we contribute to sustainable organizational success, cultivating workplaces where employees thrive, and organizations excel.</p><blockquote><em>By championing a culture of listening and empathy, we contribute to sustainable organizational success, cultivating workplaces where employees thrive, and organizations excel.</em></blockquote><p>By providing structured occasions for employees to share insights, companies can identify their needs to better address challenges and uncover opportunities to enhance workplace well-being and for organizational growth.</p><p>Here are our final practical tips for conducting successful Focus Groups:</p><ul><li><strong>Plan with clear objectives</strong>: define measurable research goals aligned with organizational needs and priorities.</li><li><strong>Select the right participants</strong>: target diverse participants with relevant backgrounds and experiences to ensure balanced perspectives.</li><li><strong>Tailor the format to your audience</strong>: use interactive activities to foster engagement, and adjust for in-person, remote, or hybrid formats.</li><li><strong>Prepare thoroughly</strong>: design clear agendas, set up logistics (tools, materials, locations), and test everything in advance to ensure smooth execution.</li><li><strong>Facilitate thoughtfully</strong>: create an inclusive environment where all voices are heard and manage group dynamics to encourage collaboration and creativity.</li><li><strong>Analyze and report</strong>: synthesize insights into actionable recommendations and collaborate with stakeholders to implement meaningful changes.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gbrAwmAzK2WyPrZjgKnwKQ.png" /><figcaption>Tips for running a Focus Group session within your company.</figcaption></figure><p>At Eni, our commitment to leveraging research methods like Focus Groups has empowered us to foster sustainable growth while cultivating a thriving workplace where employees feel valued and supported. <br>By embracing a culture of listening and action, businesses can drive meaningful changes that promote inclusivity, engagement, and overall well-being.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d67bff57deb0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/eni-digitalks/the-role-of-focus-groups-in-shaping-employee-experience-and-well-being-d67bff57deb0">The role of Focus Groups in shaping Employee Experience and well-being</a> was originally published in <a href="https://medium.com/eni-digitalks">Eni digiTALKS</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>