<?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[Stories by Zahra Shekarchi on Medium]]></title>
        <description><![CDATA[Stories by Zahra Shekarchi on Medium]]></description>
        <link>https://medium.com/@zshkarchi?source=rss-64544b506cd4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Zahra Shekarchi on Medium</title>
            <link>https://medium.com/@zshkarchi?source=rss-64544b506cd4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 03:57:48 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@zshkarchi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Leading AI Engineering: Lessons for Tech Leads]]></title>
            <link>https://medium.com/tr-labs-ml-engineering-blog/leading-ai-engineering-lessons-for-tech-leads-d7df06519add?source=rss-64544b506cd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/d7df06519add</guid>
            <category><![CDATA[technical-leadership]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <dc:creator><![CDATA[Zahra Shekarchi]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 00:25:39 GMT</pubDate>
            <atom:updated>2026-09-23T00:25:39.541Z</atom:updated>
            <content:encoded><![CDATA[<h4>Practical strategies for building reliable, production-grade AI systems.</h4><p>Moving from a successful prototype to a production-grade system demands more than technical excellence. At Thomson Reuters Labs, AI applications are developed across legal, tax, risk, and fraud, where accuracy, accountability, and reliability are strict operational requirements. Delivering AI at this standard, scale, and velocity requires a robust delivery strategy. This strategy integrates Research and Engineering along with various other elements, such as alignment with Product and stakeholders, a structured approach to technical trade-offs, and the team members involved.</p><p>This article is not an exhaustive account; it covers six lessons drawn from my experience as a tech lead in AI delivery for Practical Law, Thomson Reuters legal know-how product.</p><p>The stories behind these lessons are in my talk at <a href="https://youtu.be/P9kI903WtQU?si=q1XAFXU0iTYoSRMx">the 10th Toronto Machine Learning Summit</a>. These lessons may offer utility to any tech lead on a cross-functional team.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eYc5uJ3SL1XQ9pS8eQs6Cw.png" /></figure><h3>Alignment with Product and Stakeholders</h3><p>In my experience, a recurring delivery risk comes from a lack of agreement about what we are building toward and how (outcomes, scope, or capacity).</p><h4>1. Define “good” before kickoff</h4><p>When launching a high-priority feature, teams might plunge straight into experimentation before defining what “good” means for the first release. “Good” has to be a measurable bar on the properties users actually rely on: factual accuracy, faithfulness, and transparency. Without an explicit target, optimization runs open-ended. The team keeps improving metrics that already meet product expectations, delaying delivery.</p><blockquote><em>Takeaway: Technical and Product stakeholders must align on explicit target metrics before development begins, and track progress transparently.</em></blockquote><p><strong>Strategy:</strong> Set incremental targets instead of one distant deadline (e.g., accuracy ≥ 90% or faithfulness ≥ 95% for the first release, then commit to a 3% increase per iteration). Use tight discovery-delivery loops to track progress against these shared metrics. The transparency of these loops builds trust with Product stakeholders.</p><h4>2. Guard the team against the “hero trap”</h4><p>It is easy for tech leads to internalize delivery pressure and overestimate team capacity to accommodate stakeholders. This over-commitment forces developers to cut corners or work longer hours. Cutting corners compounds tech debt and compromises future sprint velocity. In an AI delivery context, the corners cut go beyond engineering; evaluation coverage, validation depth, and SME reviews shrink too. Longer hours leave no such trace: the sprint lands, so the over-commitment might not surface, and the inflated velocity becomes the next estimate.</p><blockquote><em>Takeaway: Guard your team against the “hero trap” with honest, data-backed estimates of capacity and development effort.</em></blockquote><p><strong>Strategy: </strong>Use empirical metrics, like historical team velocity, burn-down data, and past sprints’ completion rates, discounting periods where the team was working extra hours to hit the sprint. If resources are constrained, negotiate scope, adjust deadlines, or surface capacity limits early rather than letting the team absorb the strain.</p><h3>Technical Trade-off Decisions</h3><p>As a tech lead, I need to make and own technical decisions, with real advantages and risks. It is not always easy to know what the best option is.</p><h4>3. Treat technical debt as a calculated liability</h4><p>An effective mental model, <a href="https://c2.com/doc/oopsla92.html">coined by Ward Cunningham</a> and extended in <a href="https://martinfowler.com/bliki/TechnicalDebtQuadrant.html">Martin Fowler’s Technical Debt Quadrant</a>, is to view technical debt as a financial instrument used to buy speed-to-market. Taken intentionally, planned technical debt accelerates delivery. However, without a repayment plan, the interest compounds catastrophically. According to <a href="https://stripe.com/files/reports/the-developer-coefficient.pdf">research by Stripe</a> (2018), engineers spend roughly 17 hours per week on technical debt and maintenance. This drag creates engineering waste, opportunity cost, talent attrition, and production outage risk. That figure predates coding agents; without strict guardrails, these tools can amplify both the technical debt and its compounding interest rate.</p><blockquote><em>Takeaway: Treat technical debt as a calculated liability with an interest rate.</em></blockquote><p><strong>Strategy: </strong>To manage this liability, explicitly measure the “interest rate” (e.g., drops in velocity, outage frequency) and execute a clear management plan, such as:</p><ul><li><strong>Make debt visible:</strong> Document technical shortcuts as explicit stories in your backlog immediately.</li><li><strong>Tie reduction to business outcomes:</strong> Frame refactoring around concrete goals, such as reducing system outages or recapturing engineering hours.</li><li><strong>Practice opportunistic refactoring:</strong> Follow the “litter-pickup rule” (leave every module cleaner than you found it).</li><li><strong>Budget capacity:</strong> Dedicate a baseline percentage (e.g., 5% to 10%) of every sprint to maintenance work, adjusting this allocation based on current business risk, rather than waiting for systems to break.</li><li><strong>Onboard through tech debt:</strong> Use low-risk debt resolution as an onboarding tool, so new hires ship value while learning the codebase.</li></ul><h4>4. Leverage opposing views to de-risk architecture</h4><p>When presenting technical designs, we naturally look for validation from peers who think like us. However, the peers who push back, question assumptions, and point out flaws are your most valuable risk management resource.</p><blockquote><em>Takeaway: Teammates who disagree with your proposal are uniquely positioned to help you catch blind spots before they reach production.</em></blockquote><p><strong>Strategy: </strong>Actively invite dissent and treat objections as free architectural stress-tests, a practice that also fosters an inclusive and psychologically safe team culture. Catch blind spots early with a simple habit: pause on strong reactions, check the underlying assumptions, and pivot when the evidence supports it.</p><h4>5. Shift-left on validation feedback</h4><p>In specialized domains like legal, validating AI performance requires deep knowledge of SMEs; in our case, in-house legal experts. Thorough human-in-the-loop validation is inherently time-consuming, which means negative performance signals can arrive late and introduce delivery lag. To iterate and learn fast, we want early signals on where the system falls short.</p><blockquote><em>Takeaway: Break down evaluation into granular, high-frequency feedback cycles for fast early signals.</em></blockquote><p><strong>Strategy: </strong>Introduce rapid “spot-checks” on small, measurable samples to catch negative performance signals early. If an approach fails, you want to know immediately, so you can improve it quickly. These smaller fixes and improvements can preserve comprehensive validation as the final release gate.</p><h3>People in the Team</h3><p>While a tech lead may not formally manage team members, the way they execute their role deeply impacts everyone on the team.</p><h4>6. Quantify and recognize “glue work”</h4><p>Every project has invisible work that keeps things running smoothly. This work could be peer-unblocking, documentation, cross-team coordination, operational maintenance, and more. When left to ad-hoc distribution, this <a href="https://www.noidea.dog/glue">“glue work”</a> creates burn-out risk and single points of failure.</p><blockquote><em>Takeaway: Make invisible contributions fully visible across the team, so they can be managed more deliberately.</em></blockquote><p><strong>Strategy:</strong> Define and measure glue work for your team: tag it in the backlog or survey your team to surface it. Establish standards to capture, distribute, and perform those tasks effectively. As a result of this visibility, <strong>formally recognize and reward</strong> glue contributions.</p><h3>Conclusion</h3><p>The common theme I see in these lessons is taking something implicit and making it explicit early enough to act on. What remains implicit can only be absorbed. What’s made explicit can be negotiated, budgeted, and shared.</p><p>The increasing use of AI agents across our workflows amplifies this dynamic. Agents are highly effective at instrumentation: better logging and tracing make targets and trade-offs more visible, and backlog automation makes capacity and glue work easier to surface. Where clear practices exist, agents help make things explicit, inform decisions, and accelerate execution. Where practices and guardrails do not exist, agents compound the problem faster. Nonetheless, each of these lessons has a part that is still up to us, such as aligning on what good means, negotiating technical trade-offs, and embracing opposing views.</p><p>What do you think? Comment below, please.</p><p><em>— Views are my own.</em></p><h4><strong>References</strong></h4><ol><li>Cunningham, W. (1992). <em>The WyCash Portfolio Management System.</em> OOPSLA ’92 Experience Report. <a href="http://c2.com/doc/oopsla92.html">http://c2.com/doc/oopsla92.html</a></li><li>Fowler, M. (2009). <em>Technical Debt Quadrant.</em> <a href="https://martinfowler.com/bliki/TechnicalDebtQuadrant.html">https://martinfowler.com/bliki/TechnicalDebtQuadrant.html</a></li><li>Stripe (2018). <em>The Developer Coefficient.</em> <a href="https://stripe.com/files/reports/the-developer-coefficient.pdf">https://stripe.com/files/reports/the-developer-coefficient.pdf</a></li><li>Reilly, T. (2019). <em>Being Glue.</em> The Lead Developer. <a href="https://www.noidea.dog/glue">https://www.noidea.dog/glue</a></li><li>Shekarchi, Z. (2026). <em>Leading Trustworthy AI Engineering in Legal.</em> 10th Toronto Machine Learning Summit. <a href="https://youtu.be/P9kI903WtQU">https://youtu.be/P9kI903WtQU</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d7df06519add" width="1" height="1" alt=""><hr><p><a href="https://medium.com/tr-labs-ml-engineering-blog/leading-ai-engineering-lessons-for-tech-leads-d7df06519add">Leading AI Engineering: Lessons for Tech Leads</a> was originally published in <a href="https://medium.com/tr-labs-ml-engineering-blog">Thomson Reuters Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 2 of 3 - A Review of Fundamental AI Concepts and Applications]]></title>
            <link>https://medium.com/tr-labs-ml-engineering-blog/part-2-of-3-a-review-of-fundamental-ai-concepts-and-applications-1d7a8874f702?source=rss-64544b506cd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d7a8874f702</guid>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[tutorial]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[review]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Zahra Shekarchi]]></dc:creator>
            <pubDate>Thu, 26 Dec 2024 18:37:57 GMT</pubDate>
            <atom:updated>2025-01-02T17:57:34.848Z</atom:updated>
            <content:encoded><![CDATA[<h4>The second in a 3-part series</h4><p>In the <a href="https://medium.com/tr-labs-ml-engineering-blog/part-1-3-a-review-of-fundamental-ai-concepts-and-applications-8bf1299d5d14">previous part</a>, we reviewed the background and fundamental concepts of AI and ML. AI involves many types of data such as text (mainly natural language), vision, audio, video, and digital signals. In this part, we will zoom in on Natural Language Processing (NLP).</p><blockquote>The complete series: <a href="https://medium.com/tr-labs-ml-engineering-blog/part-1-3-a-review-of-fundamental-ai-concepts-and-applications-8bf1299d5d14">Part 1</a>, Part 2 (You are here!), <a href="https://medium.com/tr-labs-ml-engineering-blog/part-3-of-3-a-review-of-fundamental-ai-concepts-and-applications-017157b107b3">Part 3</a></blockquote><h3>NLP Fundamentals and Background</h3><p>So, let us focus on defining <strong>Natural Language Processing </strong>(NLP). This field aims to enable computers to recognize, understand, and generate human language, via text and speech, and by combining Computational Linguistic models (more rule-based) and statistical methods of ML. An NLP task is a function designed to help computers understand and process human language, such as text classification, sentiment analysis, or machine translation. In recent years, deep learning models have shown dominant performance in various NLP tasks <a href="https://www.ibm.com/topics/natural-language-processing">[1]</a>. The history of this field dates to the 1910s, and it was in the 1940s that the importance of translation became more recognized, alongside efforts to develop machines for translating messages during wars. It started by utilizing syntactic and grammatical rules and gradually, following advancements in machine learning, people made use of stochastic and probabilistic models <a href="https://cs.stanford.edu/people/eroberts/courses/soco/projects/2004-05/nlp/overview_history.html">[2]</a><a href="https://www.dataversity.net/a-brief-history-of-large-language-models/">[3]</a>. A probabilistic model is built using a statistical approach to estimate the probability of different outcomes based on data, and the model can capture the inherent uncertainty in data and incorporate it into the predictions.</p><p>A <strong>language model</strong> (LM) is a computational model of a natural or human language. It is designed to represent the language domain and can be used in different tasks such as translation or searching within a set of documents.</p><p>Some examples of first generations of language models are <strong>rule-based</strong> (used in <a href="https://en.wikipedia.org/wiki/ELIZA">Eliza</a> or <a href="https://en.wikipedia.org/wiki/SHRDLU">SHRDLU</a> chat interfaces) and <a href="https://analyticsindiamag.com/ai-mysteries/a-guide-to-hidden-markov-model-and-its-applications-in-nlp/">statistical models</a> (around 1970). Other lines of effort impacted this field (e.g., Noam Chomsky’s hierarchical <strong>grammar-based</strong> model). Later in the 1990s and 2000s, <strong>n-gram</strong> models showed a significant improvement (An n-gram is a sequence of ’n’ adjacent symbols in a particular order) <a href="https://cacm.acm.org/research/language-models/">[4]</a><a href="https://medium.com/@adria.cabello/the-evolution-of-language-models-a-journey-through-time-3179f72ae7eb">[5]</a>. In 2001, Bengio et. al. introduced the concept of word embedding using <strong>neural networks</strong> for modeling language. <strong>Word embedding</strong> represents words as vectors in a multi-dimensional space where the distance and direction between two vectors determine the similarity between the corresponding words <a href="https://www.ibm.com/topics/word-embeddings">[6]</a>. Word embedding represents a word’s meaning as a distribution, an example is illustrated below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/669/1*sv3AD28wIYegShjXX2RNyw.png" /></figure><p>Another generation of models that showed great improvements in language modeling is <strong>Recurrent Neural Networks </strong>(RNN). Although they were created in the 1980s, the introduction of GPUs and the availability of large amounts of data brought out their true potential around the 2010s. An RNN model is a deep neural network trained on <strong>sequential</strong> input data and predicting specific <strong>sequential </strong>data. For example, words are sequences of characters and sentences are sequences of words. Time series of stock prices are another example where sequential components are interrelated. In these efforts, the goal is to model longer sequences to capture more context and so better representations of word meanings <a href="https://cacm.acm.org/research/language-models/">[4]</a>.</p><p>These models still have limitations in the length of the sequence. Another significant advancement happened to alleviate RNNs’ forgetting problem by introducing an <strong>attention mechanism</strong> in an RNN by Bahdanau et. al and Sutskever et. al. in 2014. Attention is a method to determine the importance of each component in a sequence relative to the other components in that sequence. In NLP, attention is assigned to each word in a sentence. Vaswani et. al. later proposed a different implementation of the attention mechanism (self-attention <strong>Transformer architecture</strong>) in the popular paper “Attention is All You Need” and opened a new chapter of language models in 2017. The transformer model is based solely on attention mechanisms, (no recurrence and convolutions), showing superior quality in translation tasks while being more parallelizable and requiring significantly less time to train. Again, the abundance of data that can be used to train these complex models along with stronger processing units allowed for significant improvements in recent years <a href="https://cacm.acm.org/research/language-models/">[4]</a>. Language models like GPT-1, Bert, and GPT-2 are the first language models that employed and modified this architecture. Here is a timelapse of the evolution of popular language models <a href="https://levelup.gitconnected.com/the-brief-history-of-large-language-models-a-journey-from-eliza-to-gpt-4-and-google-bard-167c614af5af">[7]</a>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/593/1*RCQRIH-LVyg1KGcpYtysDQ.png" /></figure><p>To complete this part, I would like to go over the definition of the <strong>Large Language Model</strong> (LLM) we talk about daily. A large language model is an ML model pre-trained by deep learning algorithms on a vast amount of data. The underlying transformer architecture is a set of neural networks. The model learns the meanings from a sequence of text and the relationships between its words and phrases. Unlike earlier RNN-based language models, that sequentially process the input text, the transformers process entire sequences in parallel, utilizing the capabilities of GPUs and reducing the training time. In other words, this design enabled training over more and more data <a href="https://aws.amazon.com/what-is/large-language-model/">[8]</a><a href="https://www.nvidia.com/en-us/glossary/large-language-models/">[9]</a>.</p><p>The <em>intuition behind language models</em> that generate text is that, given a sequence of words or letters, they learn to predict the next element. By repeatedly making these predictions, the models can generate entire sequences, creating coherent text. These predictions are based on underlying probabilistic models designed to predict the next word in a sequence, based on patterns they learned from a vast amount of text. Because of this probabilistic nature, they might generate the next element incorrectly. This can result in errors, such as poor translations or incorrect answers from a chatbot, producing information that wasn’t in their training data and isn’t factually accurate. When the language model has a better design and is trained on a larger and higher quality dataset, the probabilistic modeling will be better, and the predictions can be more accurate. That is why we have seen better performances from the more recent LLMs.</p><p>A <strong>prompt</strong> is the input given to an LLM to generate a response. To improve the accuracy and usefulness of the responses, we can design prompts with adequate context and include example inputs and outputs <a href="https://aws.amazon.com/what-is/prompt-engineering/">[more details and techniques - 10]</a>.</p><p>It is important to mention some key <strong>limitations </strong>of LLMs as well:</p><ul><li>They are ‘large’ models, so it takes more time and data to train them, so speed and cost are significant limitations.</li><li>Accordingly, they cannot be updated frequently and retrained in real time or incrementally, so they miss new data or trends.</li><li>They can generate wrong information as well, usually with confidence (hallucination).</li><li>They struggle with tasks that require complex logical reasoning, multi-step problem-solving, or quantitative analysis.</li><li>Similar to some other AI models, biased data impacted them, so they might generate harmful biased data.</li></ul><p>A Significant portion of contributions to the AI field is to tackle these limitations <a href="https://www.textkernel.com/learn-support/blog/seven-limitations-of-llms-in-hr-tech/">[11]</a>.</p><p>It is worth talking about two techniques that resolve some of the above limitations:</p><ul><li><strong>Fine-tuning:</strong> <br>It can be considered a type of “transfer learning”, which involves utilizing the knowledge from an existing model to learn new tasks. “Pre-training”<strong> </strong>initializes model parameters to learn fundamental and general patterns from large diverse datasets. Fine-tuning takes this pre-trained model, retains the general features, and improves it using a small high-quality task-specific dataset for better performance. This technique is more cost-effective and quicker than starting from scratch <a href="https://www.ibm.com/topics/fine-tuning">[12]</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/702/1*Y_5xdgFhRob9oWmg9vV-Jg.png" /></figure><ul><li><strong>Retrieval Augmented Generation</strong>: <br>This technique improves the accuracy, quality, and reliability of data generated by trained AI models. It utilizes information from external sources that might not have been available during the training process, resulting in better-informed responses. The intuition is similar to answering questions in an open-book exam. Generally, this approach consists of an information retriever and a downstream language model. Given a user’s question, the retriever finds relevant data, and the LM uses this data to generate a response <a href="https://blogs.nvidia.com/blog/what-is-retrieval-augmented-generation/">[13]</a>.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/997/1*byQarrSWvaZg6ZGnYPUe2g.png" /></figure><p>While these techniques are helpful, they only address some issues; for instance, the model can still confidently generate inaccurate responses.</p><p>Thank you for reading this far! This part reviewed more intuitive definitions of primary terms commonly used in the NLP domain. The cited references will be useful if you want to explore further details.</p><p>The <a href="https://medium.com/tr-labs-ml-engineering-blog/part-3-of-3-a-review-of-fundamental-ai-concepts-and-applications-017157b107b3">next part</a> will review the common and popular applications developed based on these models and technologies.</p><blockquote>💬 Let the conversation begin! Was this helpful? Give it a clap! Did I get something wrong? Let me know in the comments. Or come on over to <a href="https://www.linkedin.com/groups/8168422/">Thomson Reuters Labs LinkedIn group</a> to connect and discuss further.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d7a8874f702" width="1" height="1" alt=""><hr><p><a href="https://medium.com/tr-labs-ml-engineering-blog/part-2-of-3-a-review-of-fundamental-ai-concepts-and-applications-1d7a8874f702">Part 2 of 3 - A Review of Fundamental AI Concepts and Applications</a> was originally published in <a href="https://medium.com/tr-labs-ml-engineering-blog">Thomson Reuters Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 3 of 3 - A Review of Fundamental AI Concepts and Applications]]></title>
            <link>https://medium.com/tr-labs-ml-engineering-blog/part-3-of-3-a-review-of-fundamental-ai-concepts-and-applications-017157b107b3?source=rss-64544b506cd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/017157b107b3</guid>
            <category><![CDATA[ai-applications]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Zahra Shekarchi]]></dc:creator>
            <pubDate>Thu, 26 Dec 2024 18:37:32 GMT</pubDate>
            <atom:updated>2025-01-02T17:59:38.518Z</atom:updated>
            <content:encoded><![CDATA[<h4>The third in a 3-part series</h4><p>In the previous parts, we explored fundamental concepts of AI with a particular focus on NLP. In this part, we will look into some important AI applications, especially those related to natural language.</p><blockquote>The complete series: <a href="https://medium.com/tr-labs-ml-engineering-blog/part-1-3-a-review-of-fundamental-ai-concepts-and-applications-8bf1299d5d14">Part 1</a>, <a href="https://medium.com/tr-labs-ml-engineering-blog/part-2-of-3-a-review-of-fundamental-ai-concepts-and-applications-1d7a8874f702">Part 2</a>, Part 3 (You are here!)</blockquote><h3>AI Applications</h3><p>First, we would like to define <strong>Information Retrieval </strong>(IR) and then review some of its applications. Information retrieval (IR) is finding relevant material (usually documents) of an unstructured nature (usually text) that satisfies an information need from within large data collections <a href="https://nlp.stanford.edu/IR-book/pdf/01bool.pdf">[1]</a>. So, an information retrieval system contains capabilities of the organization, storage, retrieval, and evaluation of information from information repositories <a href="https://www.geeksforgeeks.org/what-is-information-retrieval/">[2]</a>. Google or Bing search applications are over-the-web examples that retrieve relevant web pages. With the advances in AI to better understand other formats of information, we see these applications also support images, voice, or videos.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/644/1*QpntzC2uBxSFmMw78cQUaA.png" /></figure><p>AI, and particularly ML, and concepts and techniques from NLP can benefit these systems:</p><ul><li>understand and embed the data in a way that is more efficient to be retrieved <br>e.g., how to model or embed a webpage or a video so that we can compare them against a query searching a given product. What we learned about word embeddings in the <a href="https://medium.com/tr-labs-ml-engineering-blog/part-2-of-3-a-review-of-fundamental-ai-concepts-and-applications-1d7a8874f702">previous part</a> is one model.</li><li>understand the user’s intention from a query <br>If the query contains the word ‘bank’, do we mean a financial institution or the bank of a river? For instance, language models can help understand the context around the word ‘bank’ to clarify this ambiguity</li><li>train different retrieval models based on this data to provide a ranked list of results</li></ul><p>Two of the most common and frequently used applications of IR are search engines and recommendation systems.</p><ul><li>A <strong>search engine</strong> is the practical application of information retrieval techniques to large-scale data collections. It can be found in many different applications, such as web search, desktop search, or enterprise search. Search engines have been around for many years. For example, MEDLINE, the online medical literature search system, started in the 1970s. The term “search engine” initially referred to specialized hardware for text search, but gradually came to be used as the name for <em>the software system that compares queries to documents and produces ranked lists of relevant documents</em> <a href="https://ciir.cs.umass.edu/downloads/SEIRiP.pdf">[3]</a>.</li><li>A <strong>recommender system</strong> is a program that recommends products or services to users based on their predicted interest. This prediction is made based on information about the products or services, users, and interactions between them. Recommender systems aim to alleviate information overload by offering the most relevant information and services from an extensive set of options, thus delivering personalized experiences <a href="https://www.sciencedirect.com/science/article/abs/pii/S0167923615000627">[4]</a>. One of the recommendation algorithms is to recommend items with which similar users interacted previously (e.g., users from the same group of interest or engagement, age, region, or time). Another way is to recommend items similar to what a user used or engaged with before (e.g., articles from the same author or topic). This figure demonstrates the same ideas to recommend articles to users:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/1*XK-kGANMTR8JHCyLZt8tQA.png" /><figcaption>Recommending articles based on users’ engagement</figcaption></figure><p>Another AI and mainly ML application is <strong>information classification</strong>. This category has a vast range of data formats and use cases, for example,</p><ul><li>voice recognition in Amazon Alexa or Siri</li><li>spam detection in mail services</li><li>in social media or e-commerce applications, using NLP techniques, develop models to classify users’ reviews as positive and negative or classify the emotions (like happy, sad, angry, excited)</li><li>applications in medical cases of prediction, detection, diagnosis, and treatment of diseases, e.g., in cases of cancer or cognitive problems</li><li>face recognition to open our mobile phones</li></ul><p>Another category of AI applications falls under <strong>generating content</strong>, as opposed to retrieving or classifying the existing content. Now, we review some of the common generative models and applications:</p><ul><li><strong>Language translation</strong>: the conversion of text or speech from one language to another using computer algorithms</li><li><strong>Text summarization</strong>: given a set of data (usually text or speech), create a subset that represents the most important or relevant information</li><li><strong>Image or video</strong> <strong>captioning</strong>: describing or annotating the content of the image or video using a textual representation</li><li><strong>Chatbots</strong>: the application to mimic human conversation through text or voice interactions; e.g., customer service agents. As mentioned earlier, ELIZA was the first, and these applications evolved throughout time as below:</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hYcZpzYOVDNwTe_ghgMRUg.png" /><figcaption>A Short history of chatbots</figcaption></figure><ul><li>DALL·E: an AI system to create realistic images and art from a description in natural language</li><li>Assistance applications: to assist with tasks such as programming or drafting documents</li><li>These days, advanced chatbots like ChatGPT or Google Gemini are capable of a wide range of tasks including translation, summarization, captioning, grammar correction, and drafting.</li></ul><p>As AI becomes more powerful, especially in generating content, there is an increasing focus on ensuring the fairness and ethics of AI systems and their social impacts. Considering these aspects, we can efficiently utilize AI to develop innovative and responsible applications.</p><p>Thank you for staying with me up to this point. The goal of this part was to review the main applications of AI that had significant impacts on our lives as well as the products and services we use. I encourage you to review the two following pages that go over a more diverse and detailed list of AI applications:</p><ul><li><a href="https://www.simplilearn.com/tutorials/artificial-intelligence-tutorial/artificial-intelligence-applications">24 Cutting-Edge Artificial Intelligence Applications in 2024</a></li><li>And a longer list of <a href="https://research.aimultiple.com/ai-usecases/">100+ AI Use Cases &amp; Applications: In-Depth Guide for 2024</a></li></ul><blockquote>💬 Let the conversation begin! Was this helpful? Give it a clap! Did I get something wrong? Let me know in the comments. Or come on over to <a href="https://www.linkedin.com/groups/8168422/">Thomson Reuters Labs LinkedIn group</a> to connect and discuss further.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=017157b107b3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/tr-labs-ml-engineering-blog/part-3-of-3-a-review-of-fundamental-ai-concepts-and-applications-017157b107b3">Part 3 of 3 - A Review of Fundamental AI Concepts and Applications</a> was originally published in <a href="https://medium.com/tr-labs-ml-engineering-blog">Thomson Reuters Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 1/3. A Review of Fundamental AI Concepts and Applications]]></title>
            <link>https://medium.com/tr-labs-ml-engineering-blog/part-1-3-a-review-of-fundamental-ai-concepts-and-applications-8bf1299d5d14?source=rss-64544b506cd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/8bf1299d5d14</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[tutorial]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[review]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Zahra Shekarchi]]></dc:creator>
            <pubDate>Thu, 26 Dec 2024 18:36:17 GMT</pubDate>
            <atom:updated>2025-01-02T17:51:57.877Z</atom:updated>
            <content:encoded><![CDATA[<h3>Part 1 of 3 - A Review of Fundamental AI Concepts and Applications</h3><h4>The first in a 3-part series</h4><p>In this article, I will explore fundamental concepts in artificial intelligence (AI), machine learning (ML), and natural language processing (NLP), along with their common applications. The intended audience is those who do not have advanced knowledge in these domains. As AI terminology increasingly finds its way into our lives, knowing the right definitions of these concepts and techniques is important. Otherwise, we might not properly and efficiently utilize them to solve a given problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/993/1*Jj_5mvYa5O0h5-uuubSRmQ.png" /><figcaption>Generated by Microsoft Copilot</figcaption></figure><p>I recommend reading this article to get started with these topics and then learning more details from other resources such as online courses. I broke down this article into three parts: The first two outline AI and NLP fundamentals and background. The third part covers some of the main impactful applications of AI.</p><blockquote>The complete series: Part 1 (You are here!), <a href="https://medium.com/tr-labs-ml-engineering-blog/part-2-of-3-a-review-of-fundamental-ai-concepts-and-applications-1d7a8874f702">Part 2</a>, <a href="https://medium.com/tr-labs-ml-engineering-blog/part-3-of-3-a-review-of-fundamental-ai-concepts-and-applications-017157b107b3">Part 3</a></blockquote><h3>AI Fundamentals and Background</h3><p>Let us start by defining <strong>Artificial Intelligence</strong> (AI). Artificial intelligence refers to the technology that enables machines and computers to simulate human intelligence and cognitive capabilities like learning, reasoning, problem-solving, and decision-making <a href="https://www.ibm.com/topics/artificial-intelligence">[1]</a>,<a href="https://www.coursera.org/articles/what-is-artificial-intelligence">[2]</a>. In 1950, Alan Turing introduced the Turing Test to determine if a machine can exhibit intelligent behavior indistinguishable from a human. In 1956, the term Artificial Intelligence was introduced <a href="https://www.ibm.com/topics/artificial-intelligence">[1]</a><a href="https://sitn.hms.harvard.edu/flash/2017/history-artificial-intelligence/#:~:text=The%20Conference%20that%20Started%20it%20All&amp;text=It&#39;s%20considered%20by%20many%20to,and%20Marvin%20Minsky%20in%201956.">[3]</a>. Since then, there has been an ever-growing interest in developing AI applications; e.g., in technology, banking, marketing, entertainment, Bing or Google search, Apple’s Siri, self-driving cars, robotics, customer service chatbots, Google Translate, connection recommendations on Facebook or LinkedIn, and some more recent and more outstanding ones like OpenAI’s Chat GPT and Google Gemini. It is worth noting that AI is a broad field and not equivalent to machine learning (ML), deep learning (DL), and definitely not to Generative AI.</p><p>So, what is <strong>Machine Learning</strong> (ML)? It is a subfield of AI that aims to create models using algorithms trained on datasets to perform a given task, e.g., language translation or classifying facial emotions. It is fundamentally different from a “<strong>rule-based</strong>” system <a href="https://www.ibm.com/design/ai/basics/ai/#how-does-ai-work">[4]</a> in which some rules are given, and the system performs the tasks based on those rules. In contrast, ML algorithms identify patterns in the data, train the model from examples (<strong>example-based</strong>), and generalize its capability for new unseen data <a href="https://www.coursera.org/articles/what-is-machine-learning">[5]</a>. It is worth re-emphasizing that ML does not cover the entire AI landscape. AI refers to the general attempts to create machines capable of human-like abilities. Examples of AI that are not ML include the syntax-based models of human language, rule-based chatbots, <a href="https://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)">Deep Blue</a> (rule-based chess game player), <a href="https://www.geeksforgeeks.org/constraint-satisfaction-problems-csp-in-artificial-intelligence/">constraint satisfaction problem</a> (like scheduling a group meeting), and <a href="https://www.geeksforgeeks.org/minimax-algorithm-in-game-theory-set-1-introduction/">minimax search algorithm</a> (for agents playing games).</p><p>Another common term is <strong>Neural Network </strong>(NN). <strong>Artificial neural networks</strong> aim to mimic the human brain by processing input data to understand its structure through some connected units or nodes called artificial neurons (the building blocks of neural networks). A <strong>layer </strong>is the highest-level building block in a neural network that contains multiple neurons. The three types of layers of neural networks are the ‘input’ layer, the ‘hidden’ layer, and the ‘output’ layer. A ‘Deep neural network’ uses multiple hidden layers. <strong>Deep learning</strong> is a subset of ML that trains deep neural networks. These models require more training data and time while being capable of learning more complex patterns and structures <a href="https://www.ibm.com/topics/deep-learning">[6]</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/401/1*DFX3zs-uozDsPmlt4Ovuhw.png" /><figcaption>An example deep neural network</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/420/1*BdzB6-wcFyN1KJYhGjRNAQ.png" /><figcaption>This image can show the scope of domains defined above.</figcaption></figure><p><strong>Data labeling</strong> is the process of assigning target attributes, or labels, to training data. This labeled data allows ML models to learn what label predictions are expected to make. There are different approaches to training a model, especially depending on the availability of data and labels:</p><ul><li><strong>Supervised learning</strong> uses labeled datasets to train models that can predict outcomes like classes or numbers (application example: spam message detection).</li><li><strong>Unsupervised learning</strong> uses unlabeled data to discover patterns to solve clustering or association problems (application example: user segmentation for advertisement).</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/626/1*1k1UpyaBQJRnxJIX2ZJFMQ.png" /><figcaption>User Segmentation</figcaption></figure><ul><li><strong>Semi-supervised</strong> <strong>learning </strong>refers to the case in which only a part of the input data is labeled. It is beneficial when labeling data by subject matter experts is expensive <a href="https://www.ibm.com/topics/supervised-learning">[7]</a>.</li></ul><p><strong>Reinforcement</strong> <strong>learning (RL) </strong>does not belong to supervised or unsupervised learning and does not need a supervisor or a pre-labeled dataset. The idea is to reward or punish an AI agent for positive or negative behaviors in an environment until it learns the optimal behavior, like taming a lion! <a href="https://www.ibm.com/design/ai/basics/ml">[8]</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*C9yEFPQi0h8MnC2XfEU3NA.png" /><figcaption>Action-reward feedback loop of a generic RL model</figcaption></figure><p>Another key point is the difference between the <strong>parameters </strong>of an ML model and the <strong>hyper-parameters</strong>. Parameters are internal to the model and the model can learn their values from data (estimated by fitting the given input data). However, hyper-parameters are external configuration values set manually (and before any training) to determine how the model can learn. For instance, in the case of neural networks, we can configure the number of layers and the number of neurons in each layer. These are hyper-parameters. However, the values that the model will learn inside each neuron and layer (weights and biases) are not configurable. These values are the parameters and effectively what we refer to as the model.</p><p>Now we can mention another categorization of models <a href="https://machinelearningmastery.com/parametric-and-nonparametric-machine-learning-algorithms/">[9]</a>:</p><ul><li><strong>Parametric</strong>: The model keeps all the information about its predictions within a finite set of parameters, e.g., a NN.</li><li><strong>Non-parametric</strong>: The model does not make strong assumptions about the form of the mapping function, so it can learn any form of the function to map inputs and outputs. <br>In the meantime, this model can have hyper-parameters. In the example of segmenting users, the number of segments is a hyper-parameter. A decision tree is another example of a non-parametric model with a hierarchical tree structure. This category of models has grown significantly and successfully. Check <a href="https://www.coursera.org/in/articles/decision-tree-machine-learning">this link</a> to learn more about decision trees.</li></ul><p>Another categorization of learning algorithms in ML is based on how we define the underlying hypothesis:</p><ul><li><strong>Discriminative</strong>: The algorithm learns some boundaries to separate (discriminate) data points into different classes, e.g., to assign an email to spam and non-spam categories.</li><li><strong>Generative</strong>: The algorithm learns the actual data distribution and different data points to generate new data points, instead of the boundaries. For instance, when we send messages, the predictive text feature generates and suggests the next words.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M8grVT2sDcwTsEMyVCXTYg.png" /><figcaption>The discriminative model draws boundaries, while the generative model learns how to describe and generate data.</figcaption></figure><p>Thank you for your attention up to this point. I hope you found the information helpful! There is much more to learn about the AI fundamentals and more than one page is needed. The cited links can be one way to dive deeper into the details of the above concepts. We move to the <a href="https://medium.com/tr-labs-ml-engineering-blog/part-2-of-3-a-review-of-fundamental-ai-concepts-and-applications-1d7a8874f702">next part</a> to focus on one of the rapidly evolving subdomains of AI, Natural Language Processing.</p><blockquote>💬 Let the conversation begin! Was this helpful? Give it a clap! Did I get something wrong? Let me know in the comments. Or come on over to <a href="https://www.linkedin.com/groups/8168422/">Thomson Reuters Labs LinkedIn group</a> to connect and discuss further.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8bf1299d5d14" width="1" height="1" alt=""><hr><p><a href="https://medium.com/tr-labs-ml-engineering-blog/part-1-3-a-review-of-fundamental-ai-concepts-and-applications-8bf1299d5d14">Part 1/3. A Review of Fundamental AI Concepts and Applications</a> was originally published in <a href="https://medium.com/tr-labs-ml-engineering-blog">Thomson Reuters Labs</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>