<?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 nicolette mtisi on Medium]]></title>
        <description><![CDATA[Stories by nicolette mtisi on Medium]]></description>
        <link>https://medium.com/@nicmtisi?source=rss-2195b4c1bcfc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ZWQtZo8KQhMWAbtbrVRHKQ@2x.jpeg</url>
            <title>Stories by nicolette mtisi on Medium</title>
            <link>https://medium.com/@nicmtisi?source=rss-2195b4c1bcfc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 09:35:17 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@nicmtisi/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[I Built a Healthcare Claims Data Pipeline to Practice the Way Real Data Teams Actually Work]]></title>
            <link>https://medium.com/@nicmtisi/i-built-a-healthcare-claims-data-pipeline-to-practice-the-way-real-data-teams-actually-work-2581f0c5cfda?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/2581f0c5cfda</guid>
            <category><![CDATA[sql]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Thu, 10 Sep 2026 21:55:54 GMT</pubDate>
            <atom:updated>2026-09-10T21:55:54.314Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qJoyFvceC8I_giK9pFDqPA.png" /></figure><p>A few weeks ago, dbt was just an acronym I kept seeing in job postings and skimming past. I’d read “experience with dbt” on a dozen data analyst and analytics engineer listings without a clear picture of what that experience should look like. So I decided to stop guessing and build something with it.</p><p>Here’s what I built, and why I think it’s a useful project, not just a checkbox.</p><h3>The problem I wanted to solve</h3><p>Every time someone gets billed for a doctor’s visit, a procedure, or a scan, that generates a claim. Insurers and researchers use this data to answer real questions: is this procedure priced fairly? Does the same treatment cost more in one region than another? Those questions only get answered if the underlying data is clean, consistent, and something people can actually trust.</p><p>That last part is the piece most people skip when they talk about data work. Cleaning data once, in a notebook, as a one-off script, is easy. Building something that produces the <em>same trustworthy result every single time you run it</em>, with automatic checks in place to catch problems before they reach a dashboard, is a different skill entirely. That’s the job of an analytics engineer, and that’s what I wanted to practice.</p><h3>What dbt actually does</h3><p>I want to be precise about this, because it’s easy to describe dbt wrong. dbt doesn’t clean data in real time as it comes in. It’s a batch transformation tool: you point it at data that already exists, and it applies your SQL logic to reshape it, on a schedule or on demand. Real-time cleaning is a different category of tool entirely (think streaming pipelines).</p><p>What dbt is genuinely good at is <strong>repeatability and trust</strong>. You write your transformation logic once, in plain SQL, and every time you run it, dbt applies it the same way, documents what it did, and tests the result automatically. Anyone on a team can open your project and know exactly what happened to the data and why, instead of trusting someone’s memory of a script they ran once six months ago.</p><p>That’s the actual value proposition, and it’s the one I built this project around.</p><h3>What I built</h3><p>I modeled the project on the structure of CMS’s public Medicare payment data, and built a full pipeline in three stages.</p><p><strong>Raw data.</strong> Provider records, procedure codes, and claim-level payment data, including the kind of mess real data has: duplicate rows, missing values, inconsistent formatting.</p><p><strong>Cleaning and standardizing.</strong> Using dbt, I wrote SQL to strip stray whitespace from text fields, force values into the correct data types, fill in reasonable estimates for missing values, and remove genuine duplicate rows (not just similar-looking ones, but rows where every value matched exactly).</p><p><strong>Modeling into a star schema.</strong> This is the core of the project. I built one fact table, fact_claims, holding the actual events and numbers: how many claims, how much was paid. I surrounded it with four dimension tables describing who, what, where, and when: dim_provider, dim_procedure, dim_geography, dim_date. Every dimension table has a clean, consistent ID column, and the fact table connects to each one through joins, so anyone querying this model can ask real business questions without wading through raw, disconnected tables.</p><p><strong>Testing.</strong> I wrote 17 automated data quality tests: checks that IDs are unique and never blank, checks that every ID referenced in the fact table actually exists in its corresponding dimension table, and a custom check confirming no payment amount is ever negative or zero. All 17 pass, which means the model holds together the way it’s supposed to, and any future changes to the data would immediately surface if they broke that consistency.</p><p><strong>Output.</strong> The final result is a clean, dashboard-ready table. I built a chart in Looker Studio showing average payment amount by procedure, broken out by region, which surfaces a clear pattern worth investigating further.</p><h3>One honest note</h3><p>The data in this project is synthetic, generated to match CMS’s real schema, since I didn’t have access to production claims data. The pipeline logic and structure are identical to what you’d build against the real dataset; only the underlying numbers are simulated.</p><h3>Why I think this is a useful project</h3><p>It’s not a research finding, and it was never meant to be one. Its value is that it demonstrates a real, repeatable workflow: take messy raw data and turn it into something reliable, tested, and query-ready, the same way data teams actually do it. That’s a concrete, defensible skill, and one I can now walk through in detail, from the cleaning logic to the schema design to why each test exists.</p><p>The full project, including the code, tests, and dashboard, is available on my <a href="https://github.com/nic-stack/healthcare-claims-pipeline">GitHub</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2581f0c5cfda" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When Your Product Makes Your Customer Look Good, They Never Leave]]></title>
            <link>https://medium.com/underrwritten/when-your-product-makes-your-customer-look-good-they-never-leave-026d72711501?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/026d72711501</guid>
            <category><![CDATA[business]]></category>
            <category><![CDATA[data-analytics]]></category>
            <category><![CDATA[insuretech]]></category>
            <category><![CDATA[insurance]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Tue, 21 Apr 2026 01:36:01 GMT</pubDate>
            <atom:updated>2026-04-24T01:02:33.217Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The insurance industry processes trillions of dollars in claims every year. Most of it moves like it is 1995.</em></p><p>Imagine this.</p><p>Your car was just totaled. You did not see the other driver coming. You are standing on the side of the road, shaken, calling your insurance company. The agent is polite. The adjuster is efficient. The claim is approved within 48 hours. You feel relieved.</p><p>Then you wait.</p><p>Three days pass. Then five. Then ten. You call the company. They tell you the check is in the mail. On day fourteen, it arrives. A paper check. You have to deposit it, wait for it to clear, and then finally, two and a half weeks after the worst day of your month, you have access to your own money.</p><p>By the time that check clears, the relief you felt on day two is long gone. What remains is frustration. And at renewal time, when your insurer asks if you want to stay, you remember the wait.</p><p>This is not a hypothetical. This is the average claims experience in the United States today.</p><h3>The Data Behind the Frustration</h3><p>The numbers are striking when you look at them together.</p><p>According to the <a href="https://www.jdpower.com/business/press-releases/2024-us-auto-claims-satisfaction-study">J.D. Power 2024 U.S. Auto Claims Satisfaction Study</a>, the average repair cycle time for a claim is 22.3 days. The average cost to process a single claim sits between $15 and $25, driven largely by manual workflows, paper-based disbursements, and human error correction. The average payment error rate across the industry sits at approximately 3%, meaning 3 out of every 100 claim payments require reissuance, adjustment, or reversal after the fact.</p><p>These are not small inefficiencies. For a carrier processing 50,000 claims per year, a 3% error rate means 1,500 payments that go wrong. At $20 per claim in processing cost, that is $1 million annually in operational overhead, before accounting for the cost of frustrated policyholders and the staff hours required to fix mistakes.</p><p>And then there is the less visible cost: the policyholder who does not renew.</p><p>The <a href="https://www.jdpower.com/business/insurance/auto-claims-satisfaction-study">same J.D. Power research</a> found that 80% of auto insurance customers who have poor claims experiences have already left or say they plan to leave their carrier. Claims satisfaction is one of the strongest predictors of customer retention in insurance. Carriers with high claims satisfaction scores retain significantly more policyholders at renewal. Carriers with low scores, driven largely by payment speed and payment options, lose customers at a rate that compounds year over year.</p><p>The payment is not just a transaction. It is the moment of truth.</p><h3>Why Payments Are Still Stuck</h3><p>If the problem is so clear, why has the industry been slow to fix it?</p><p>The honest answer is that insurance is one of the most complex, regulated, and legacy-dependent industries in the world. Most carriers still operate on core systems built decades ago. Replacing payment infrastructure means touching compliance workflows, state regulatory requirements, vendor contracts, and internal processes that have been in place for years. The incentive to move fast has historically been low because switching costs are high and most carriers assumed their competitors were equally slow.</p><p>They were right, until they were not.</p><p>The rise of insurtech has introduced a new class of payment infrastructure built specifically for insurance. <a href="https://www.mckinsey.com/industries/financial-services/our-insights/claims-2030-dream-or-reality">McKinsey’s research on claims modernization</a> shows that insurers who implement integrated digital claims solutions experience up to 30% reduction in processing time and significant operational cost savings. These platforms do not require carriers to overhaul their entire technology stack. They integrate with existing systems and begin processing digital payments within weeks rather than months. The integration timeline that used to take 24 weeks now takes 2 to 4.</p><p>The excuse of complexity is shrinking. The carriers who are still waiting are now the ones falling behind.</p><h3>The Retention Argument Nobody Is Making Loudly Enough</h3><p>Here is the insight that does not get enough attention in conversations about insurtech payments:</p><blockquote>A payment company is not back-office infrastructure. It is a front-line retention tool.</blockquote><p>As someone working inside the insurtech payments space at <a href="https://snaprefund.io/">SnapRefund</a>, I have seen this dynamic up close. The carriers who are most excited about modernizing their payment infrastructure are not primarily excited about cost savings, though those are real and significant. They are excited about what faster, better payments do for their relationship with their policyholders.</p><p>When a claimant files a claim and receives their money within 48 hours, digitally, in whatever payment method they prefer, they do not think about the technology that made it happen. They think about their carrier. They think: that was easy. That was fast. That company came through for me when I needed them.</p><p>That feeling is worth more than any marketing campaign.</p><p><a href="https://www.deloitte.com/us/en/insights/industry/financial-services/insurance-claims-transformation.html">Deloitte’s Future of Claims research</a> makes this point directly: claims operations, which have been traditionally treated as outputs of a reactive back office, must become a powerful differentiator. The research notes that 67% of policyholders cite claims handling efficiency as the primary factor in their decision to renew policies. That is not a marginal metric. That is the retention story in a single number.</p><p>And here is the principle that I keep coming back to:</p><blockquote>When your product makes your customer look good to their customers, your customer never wants to leave you.</blockquote><p>A payment platform that makes a carrier look fast, reliable, and modern to their policyholders is not a vendor. It is a competitive advantage embedded in the carrier’s most important customer touchpoint. The carrier that pays instantly does not just win on speed. They win on trust. They win on loyalty. They win at renewal.</p><h3>What Instant Payment Actually Changes</h3><p>Speed is the headline but it is not the whole story. Here is what modernized claims payment infrastructure actually delivers:</p><p><strong>For the claimant </strong>No paper check to wait for. No vendor account to set up. No app to download. Payment arrives in the method they already use, whether that is a bank transfer, digital wallet, or whatever they prefer. The friction that used to define the claims experience disappears.</p><p><strong>For the carrier </strong>Cost per claim drops significantly. Error rates fall because digital payments create a full audit trail that catches mistakes before they become disbursements. The number of payment-related support calls decreases. Compliance becomes easier to demonstrate because every transaction is logged and traceable. And claimant satisfaction scores, the metric that most directly predicts renewal, improve.</p><p><strong>For the carrier’s bottom line </strong><a href="https://www.mckinsey.com/industries/financial-services/our-insights/insurance-blog/will-your-insurance-it-investments-pay-off">McKinsey’s analysis of insurance IT modernization</a> found that carriers deploying modernized claims technologies can improve profit margins by 25 to 40%. The combined ratio, which is the insurance industry’s primary measure of profitability, improves on both sides. Fewer errors and faster settlement reduce the claims cost component. Lower operational overhead reduces the expense component. A 1 to 2 point combined ratio improvement, which modernized payment infrastructure makes achievable, is worth millions at scale.</p><p><strong>For the relationship </strong>This is the one that matters most. Every carrier is selling roughly the same product, a promise to pay when something goes wrong. The carrier that fulfills that promise fastest and most smoothly wins the relationship. Payment speed is not a feature. It is the product.</p><h3>The Carriers Who Get This Are Already Moving</h3><p>The conversation in the industry has shifted. <a href="https://www.deloitte.com/us/en/insights/industry/financial-services/financial-services-industry-outlooks/insurance-industry-outlook.html">Deloitte’s 2026 Global Insurance Outlook</a> notes that customer experience is no longer a peripheral metric. It drives both retention and growth. The report highlights that today’s policyholders, especially in the P&amp;C segment, want speed and hyper-personalized solutions delivered seamlessly across digital and human touchpoints.</p><p>New MGAs building on modern technology stacks are adopting digital payment rails from day one. They do not have legacy systems to work around. They are setting a new standard that incumbent carriers will eventually have to match.</p><p>The more forward-looking carriers are already in motion. They are not waiting for regulation to force the change. They are recognizing that every day they process payments slowly is a day their policyholder’s experience is worse than it needs to be, and a day their competitor has an opportunity to look better.</p><p>The ones still mailing paper checks are not just running an inefficient process. They are actively making their product worse at the moment that matters most.</p><h3>The Principle That Changes How You Think About This</h3><p>I want to close with the idea I started with, because I think it deserves to be said plainly.</p><p>In every industry, the companies that build lasting customer relationships are the ones that make their customers successful. Not just their direct customers. Their customers’ customers.</p><p>In insurance, the carrier is the customer. The policyholder is the customer’s customer. A payment platform that makes the carrier look fast, competent, and caring to their policyholders is not just selling a service. It is making the carrier better at the thing that matters most in their business: being there when it counts.</p><blockquote>When your product makes your customer look good to their customers, your customer never wants to leave you.</blockquote><p>That is not a tagline. It is the business case for why payment speed in insurance is not an operational detail. It is a strategic imperative.</p><p>The carriers who understand this are building moats. The ones who do not are building frustration, one paper check at a time.</p><p><em>Nicolette Mtisi is a data analyst and independent researcher covering the business of insurance and insurtech. She is completing her MS in Data Analytics at Yeshiva University’s Katz School of Science and Health and works in the insurtech payments space. She writes at Underwritten, a publication focused on independent analysis of the insurance industry.</em></p><p><strong>Sources</strong></p><p>1. <a href="https://www.jdpower.com/business/press-releases/2024-us-auto-claims-satisfaction-study">J.D. Power 2024 U.S. Auto Claims Satisfaction Study</a> | J.D. Power, October 2024</p><p>2. <a href="https://www.mckinsey.com/industries/financial-services/our-insights/claims-2030-dream-or-reality">Claims 2030: Dream or Reality</a> | McKinsey &amp; Company</p><p>3. <a href="https://www.mckinsey.com/industries/financial-services/our-insights/insurance-blog/will-your-insurance-it-investments-pay-off">Will Your Insurance IT Investments Pay Off</a> | McKinsey &amp; Company</p><p>4. <a href="https://www.deloitte.com/us/en/insights/industry/financial-services/insurance-claims-transformation.html">Preserving the Human Touch in Insurance Claims Transformations</a> | Deloitte Insights</p><p>5. <a href="https://www.deloitte.com/us/en/insights/industry/financial-services/financial-services-industry-outlooks/insurance-industry-outlook.html">2026 Global Insurance Outlook</a> | Deloitte Insights</p><p>6. <a href="https://www.jdpower.com/business/insurance/auto-claims-satisfaction-study">J.D. Power U.S. Auto Claims Satisfaction Study Overview</a> | J.D. Power</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=026d72711501" width="1" height="1" alt=""><hr><p><a href="https://medium.com/underrwritten/when-your-product-makes-your-customer-look-good-they-never-leave-026d72711501">When Your Product Makes Your Customer Look Good, They Never Leave</a> was originally published in <a href="https://medium.com/underrwritten">Underrwritten</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Algorithm That Learns How to Learn: A Student's Guide to Evolutionary AI]]></title>
            <link>https://medium.com/@nicmtisi/the-algorithm-that-learns-how-to-learn-a-students-guide-to-evolutionary-ai-0847f38377a3?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/0847f38377a3</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 19:22:56 GMT</pubDate>
            <atom:updated>2026-04-14T19:22:56.370Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hMgw7wiHIlCQZYLUy9fVow.jpeg" /><figcaption>AI Evolution</figcaption></figure><h3>The Algorithm That Learns How to Learn: A Student’s Guide to Evolutionary AI</h3><p><em>By Nicolette Mtisi | M.S. Data Analytics &amp; Visualization, Yeshiva University</em></p><p>I am not a computer scientist by training. I came into my graduate program as a writer, a strategist, someone who works with words and ideas. So when my AGI course handed me an assignment on evolutionary algorithms and LLM-guided program search, I will be honest. I stared at the screen for a long time.</p><p>But somewhere between implementing a genetic algorithm from scratch and watching an AI use Google Gemini to rewrite its own broken code, something clicked. This was not just an assignment. This was a window into how the next generation of AI systems actually works.</p><p>If you are a student navigating a technical degree, especially one that did not start in computer science, this article is for you.</p><h3>The Big Idea: What If You Let Solutions Evolve?</h3><p>The central question of this assignment was deceptively simple:</p><p><em>Instead of programming a solution by hand, what if you defined what “good” looks like and let an algorithm discover the answer on its own?</em></p><p>This is the core idea behind evolutionary algorithms. You start with a population of candidate solutions. Random, imperfect, maybe completely wrong. You measure how good each one is (its “fitness”). You keep the best ones, combine them, introduce small random changes, and repeat. Generation by generation, your population improves.</p><p>It sounds almost biological, because it is. The entire framework is inspired by natural selection. And it turns out that this simple idea, select, combine, mutate, repeat, is powerful enough to solve problems that traditional programming cannot.</p><h3>Part 1: The Genetic Algorithm — Learning to Search Intelligently</h3><p>The first thing I built was a classical genetic algorithm (GA) to solve a simple optimisation problem: find a vector of ten numbers that are all as close to zero as possible.</p><p>Three operators made this work:</p><p><strong>Tournament Selection</strong> — instead of picking parents randomly, you run a mini competition. Grab a few individuals from the population, pick whoever scored highest. Good solutions get to reproduce. Bad ones mostly do not.</p><p><strong>Crossover</strong> — two parents produce a child by mixing their genes. For each position in the vector, flip a coin. Heads takes from parent A, tails from parent B. The child inherits strengths from both.</p><p><strong>Mutation</strong> — after crossover, add a tiny random wobble to every gene. This keeps diversity in the population so you do not converge too early on a local optimum.</p><p>The result? After 100 generations the algorithm found solutions dramatically better than pure random search using the same number of evaluations. This was my first real lesson:</p><p><strong>Intelligent search that builds on good solutions beats blind guessing every time.</strong></p><p>This is directly applicable to data science. Hyperparameter tuning, finding the right learning rate, tree depth, number of layers for a neural network, is the same problem. You are searching for the best combination of values. Companies like Google use evolutionary algorithms to do this automatically in their AutoML systems.</p><h3>Part 2: MAP-Elites — When One Best Answer Is Not Enough</h3><p>The genetic algorithm finds one winner. But what if you needed the best solution for every possible type of situation?</p><p>This is the question <a href="https://arxiv.org/abs/1504.04909">MAP-Elites</a> (Multi-dimensional Archive of Phenotypic Elites) was designed to answer. Instead of maintaining a single population converging toward one optimum, MAP-Elites maintains a grid archive. Each cell in the grid represents a different “type” of solution (called a behaviour descriptor), and each cell stores the best solution found for that type.</p><p>In my implementation, the grid was 20x20, giving 400 cells, each representing a different region of the solution space. After 1000 iterations, MAP-Elites filled 69% of the grid with high-quality, diverse solutions. A regular GA using the same budget filled only 9%.</p><p>The trade-off is real and important:</p><p>Goal Better Method Find the single best solution Genetic Algorithm Find the best solution for every type MAP-Elites Understand the full landscape of solutions MAP-Elites Drug discovery, robotics, game AI MAP-Elites Speed and simplicity Genetic Algorithm</p><p>This distinction matters enormously in practice. In robotics, MAP-Elites is used to evolve robots that can keep walking even after losing a leg, because the archive contains good locomotion strategies for every possible body configuration. In drug discovery, you do not want one perfect drug molecule. You want the best molecule for every type of patient.</p><p><strong>Lesson: diversity in solutions is often more valuable than a single optimal answer.</strong></p><h3>Part 3: LLM-Guided Mutation — When AI Becomes the Mutation Operator</h3><p>This is where the assignment became genuinely exciting.</p><p>In Parts 1 and 2, individuals were vectors of numbers. In Part 3, individuals became actual Python programs, source code stored as strings. Fitness was no longer how close to zero, but how many test cases your function passed.</p><p>The target task: evolve a Python function that correctly sorts a list.</p><p>We started with broken seed programs:</p><pre>def solve(lst):<br>    return []  # fitness: 0.10, always returns empty</pre><pre>def solve(lst):<br>    return lst  # fitness: 0.40, returns unsorted</pre><p>Two mutation strategies were compared:</p><p><strong>Random mutation</strong> shuffles lines of code randomly. Like fixing an essay by cutting it up and rearranging sentences. Usually makes things worse.</p><p><strong>Gemini-guided mutation</strong> sends the broken program to Google Gemini along with a detailed failure report. Gemini reads which test cases failed and why, then rewrites the function to fix them.</p><p>The difference was stark. In my actual results: random mutation improved fitness <strong>0% of the time</strong>. Gemini-guided mutation improved fitness <strong>100% of the time</strong>.</p><p>This is not a gimmick. This is the architecture behind <a href="https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/">FunSearch</a>, DeepMind’s 2023 system that used LLMs as mutation operators to discover new mathematical algorithms, solving problems that had stumped mathematicians for decades. The same architecture, in principle, as what I implemented in a Google Colab notebook.</p><p><strong>Lesson: LLMs are not just chatbots. They are semantic mutation operators that understand the structure and intent of code.</strong></p><h3>Part 4: UCB1 — Learning Which Tool to Use</h3><p>The final piece was an elegant solution to a practical problem: when you have multiple mutation operators available, how do you decide which one to use?</p><p>UCB1 (Upper Confidence Bound) is a classic algorithm from the multi-armed bandit literature. It tracks the average reward of each operator but adds an exploration bonus for operators that have not been tried much. The formula:</p><pre>UCB(k) = mean_reward + c x sqrt(ln(total_tries) / tries_for_operator)</pre><p>The bonus gets smaller the more you use an operator. This forces the algorithm to try everything before settling on what works. Over 300 rounds, UCB1 correctly identified the best operator (Guided, with true success rate 0.5) and selected it <strong>147 times</strong> versus only <strong>42 times</strong> for the worst operator (Random, with true success rate 0.2). Nobody told it upfront which was better. It figured that out entirely on its own.</p><h3>The Full Pipeline — Watching AI Write Its Own Code</h3><p>This is where everything came together. The pipeline combined the archive, UCB operator selection, and Gemini mutation into one loop:</p><pre>select parent → UCB picks operator → mutate → validate syntax → test → insert into archive → repeat</pre><p>The results were genuinely astonishing. Starting from broken seed programs with fitness of 0.70, the pipeline evolved a <strong>perfect sorting function in just 5 iterations</strong>:</p><pre>Generation 0  → fitness 0.70  (broken bubble sort)<br>Generation 5  → fitness 1.00  (def solve(lst): return sorted(lst))<br>Generation 20 → fitness 1.00  (still perfect!)</pre><p>Final best fitness: <strong>1.00</strong>, passing all 10 test cases perfectly. The system figured out that return sorted(lst) was the answer without anyone telling it. It got there by failing, learning from failure, and trying smarter mutations each time.</p><p><strong>Lesson: when you combine smart search (UCB) with smart mutation (Gemini), the system learns both WHAT to search for and HOW to search for it.</strong></p><h3>The Limits — What This Approach Cannot Do</h3><p>Intellectual honesty requires acknowledging where evolutionary program search breaks down.</p><p><strong>Scalability</strong> is a real constraint. Evolving a ten-line sorting function is tractable. Evolving a production codebase with thousands of functions would require astronomical computational resources. The search space grows exponentially with program complexity.</p><p><strong>Representation matters enormously.</strong> Treating programs as flat strings is crude. A one-character mutation can break everything or produce something semantically unrelated. More sophisticated representations, like abstract syntax trees or modular program structures, would make mutations more meaningful.</p><p><strong>The fitness landscape can be deceptive.</strong> Most random mutations score zero. There are vast empty regions where the algorithm receives no learning signal. Local optima like a partially working sort function can trap search far from globally optimal solutions.</p><p><strong>LLMs hallucinate.</strong> Gemini can confidently produce code that looks correct but fails edge cases. It can overfit to the specific test cases provided and hardcode special cases rather than learning the general algorithm. The LLM’s understanding is probabilistic, not guaranteed.</p><p><strong>Safety is non-trivial.</strong> Autonomously generating and executing code in an unsandboxed environment is genuinely dangerous. Evolved programs run with the same system permissions as their host environment. This is not theoretical. It is a real constraint that production systems like FunSearch had to engineer around carefully.</p><h3>What This Means for Students Like Us</h3><p>If you are in a data analytics, data science, or AI program right now, here is what I want you to take away:</p><p><strong>1. The boundaries between fields are dissolving.</strong> Evolutionary computation, LLMs, reinforcement learning, and program synthesis are converging into systems that can discover knowledge autonomously. Understanding any one of these deeply gives you leverage across all of them.</p><p><strong>2. AutoML is built on these foundations.</strong> When you use automated hyperparameter tuning in scikit-learn, or Neural Architecture Search in deep learning frameworks, you are using the descendants of the algorithms covered here. Understanding the foundations makes you a better practitioner.</p><p><strong>3. The most interesting AI applications define a fitness function and let the system search.</strong> You do not always need to specify the solution. You need to specify what a good solution looks like. This is a fundamentally different way of thinking about problem-solving.</p><p><strong>4. Code literacy is not optional anymore.</strong> Even if you came into tech through a non-traditional path, the ability to read, understand, and reason about code is becoming as fundamental as reading and writing. This work pushed me to engage with Python at a deeper level than I expected, and I am a better analyst for it.</p><h3>Closing Thoughts</h3><p>I started this work confused. I finished it genuinely inspired.</p><p>Not because the code was easy. It was not. Not because everything ran perfectly. It did not. I spent an embarrassing amount of time fighting the Gemini API, trying three different model names before finding one that worked. But those debugging moments taught me something too. Real engineering is messy, APIs change, models get deprecated, and the ability to problem solve in the moment matters as much as knowing the theory.</p><p>What inspired me was understanding, by the end, that I had built in miniature the same architecture that DeepMind used to make mathematical discoveries. Starting from broken code with a fitness of 0.70, my pipeline evolved a perfect sorting function in 5 iterations. Final best fitness: 1.00. I had touched something real.</p><p>If you are a student feeling overwhelmed by the technical depth of an AI or data science program, the concepts are learnable. The gap between where you are and where these ideas live is smaller than it looks. Start with the fundamentals. Ask why, not just how. And keep building.</p><p>The algorithms evolved. So can we.</p><h3>Further Reading</h3><ul><li><a href="https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/">FunSearch blog post</a> — DeepMind’s accessible explanation, start here</li><li><a href="https://www.nature.com/articles/s41586-023-06924-6">FunSearch paper</a> — “Mathematical discoveries from program search with large language models”, Nature 2023</li><li><a href="https://arxiv.org/abs/1504.04909">MAP-Elites paper</a> — “Illuminating search spaces by mapping elites”, Mouret and Clune 2015</li><li><a href="https://lilianweng.github.io/posts/2019-09-05-evolution-strategies/">Evolution Strategies</a> — Lilian Weng’s clear breakdown of evolutionary methods in AI</li><li><a href="https://github.com/google-deepmind/funsearch">FunSearch on GitHub</a> — the actual open source code from DeepMind</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0847f38377a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Was “Good With Money” and Still Broke: My Personal Money Code as a Student]]></title>
            <link>https://medium.com/@nicmtisi/i-was-good-with-money-and-still-broke-my-personal-money-code-as-a-student-c502d662b57d?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/c502d662b57d</guid>
            <category><![CDATA[budgeting]]></category>
            <category><![CDATA[students]]></category>
            <category><![CDATA[personal-finance]]></category>
            <category><![CDATA[financial-independence]]></category>
            <category><![CDATA[young-professionals]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Fri, 23 Jan 2026 14:02:05 GMT</pubDate>
            <atom:updated>2026-01-23T14:02:05.451Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A framework for managing money as a student or recent graduate when traditional budgeting advice does not fit.</em></p><p>I used to think I was good with money because I paid my credit card on time. I tracked my spending. I hit “pay in full” every month. I did everything the internet says responsible people should do.</p><p>And I was still broke.</p><p>Not technically broke. I was not overdrafting or missing payments. But I was stuck in a strange middle ground where I was doing all the “right things” and still felt one unexpected expense away from disaster.</p><p>The problem was not discipline.<br> It was clarity.</p><p>In 2025, I entered a completely new environment as an international student in the United States. Everything was unfamiliar. The cost of living. The way people spent money. The expectations around lifestyle. I did not know what was normal, what was necessary, or what was simply me trying to fit in.</p><p>Credit cards were a big part of that confusion. There were no credit cards in my country. This was a completely new money experience, and one that can trap you quickly if you are not careful.</p><p>By the end of 2025, I realized I did not have a framework for what money was actually for in this specific season of my life. I was following generic advice designed for people with margin. People who had already covered rent, loans, and groceries and still had money left over.</p><p>That was not me.</p><p>And if you are a student or a recent graduate, especially navigating a new country or city, it probably is not you either.</p><p>So in 2026, I made a conscious decision to use money wisely. Not perfectly. Not restrictively. But intentionally, with a framework that actually fits the season I am in.</p><p>I wrote myself a money code. Not rules to restrict me, but principles to protect me.<br> Here is what I live by now.</p><h3>1. Every dollar gets a job</h3><p>Unassigned money is not extra money.</p><p>If money is left over after rent, bills, and groceries, it does not just sit there waiting to be spent. It gets assigned immediately.</p><p>I divide my money into clear categories. Bills cover rent, utilities, and loans. Emergency savings cover medical bills, a broken laptop, or anything urgent and unexpected. Dream savings are for future goals that are not urgent but still matter to me. The spending account is what I can actually spend without guilt.</p><p>Money without a job will find one. And it is usually not the job you needed it to do.</p><p>Assigning money before I need it removes the mental negotiation every time I want to buy something. I am no longer asking, “Can I afford this?” I am asking, “Does my spending account cover this?”</p><p>If the answer is yes, I spend it.<br> If the answer is no, it waits.</p><h3>2. Paying it off does not mean I could afford it</h3><p>Just because I can clear my credit card does not mean the purchase was sustainable.</p><p>This was a major realization for me. I would spend on my card, pay it off when my paycheck came in, and feel responsible. But I was not paying attention to where the repayment money was coming from.</p><p>If clearing my card meant pulling from rent money, bill money, or loan money, then the purchase was too expensive for this season. Period.</p><p>Paying something off and affording something are not the same thing.</p><h3>3. Credit cards require margin, not discipline</h3><p>I only use credit when I already have the cash to cover it.</p><p>Credit cards are often marketed as convenience tools, but they only work as intended when you have margin. Without margin, they become a timing game. You end up borrowing from future paychecks that are already spoken for.</p><p>I do not use my card to make it to payday. I use it when the cash is already sitting in my account and I want the points or the fraud protection.</p><p>Credit is a tool, not a bridge between paychecks.</p><h3>4. I do not reward discipline with spending</h3><p>Skipping plans does not create permission to spend elsewhere.</p><p>I used to do this often. I would skip going out with friends to save money, then reward myself by ordering takeout or buying something online. The money still left. I just redirected the guilt.</p><p>Now I recognize this for what it is. Emotional spending dressed up as a treat.</p><p>Security is the reward. Stability is the treat. Saying no to one thing does not unlock budget for another unless it was planned.</p><h3>5. “I deserve it” is not a budget category</h3><p>Emotions do not decide spending.</p><p>I am not saying I never spend on things I want. I am saying I do not let feelings override the plan in the moment.</p><p>If it is not already accounted for, it is a no. Even if I want it. Even if I have had a hard week. Even if everyone else is doing it.</p><p>Deserve has nothing to do with it. The question is whether it fits the plan or compromises something more important.</p><h3>6. If there is no money allocated, the answer is no</h3><p>When an expense comes up and there is no money assigned to that category, I do not have the money. Period.</p><p>I do not borrow from another account hoping to recover it later. That is how you end up living in circles, constantly robbing one bucket to fill another and never actually getting ahead.</p><p>If my spending account is empty and something comes up, the answer is no unless it is a genuine emergency. Wanting it badly does not make it an emergency.</p><p>This also means paying attention to small expenses. Ten dollars feels insignificant in the moment. But ten dollars spent daily without thinking turns into three hundred dollars by the end of the month. That is rent money. That is bill money. That is the difference between stability and stress.</p><p>Unbudgeted spending adds up faster than most people realize.</p><h3>7. Unspent money is future peace</h3><p>Saving is not boring. It is protection.</p><p>I used to think saving was about delayed gratification. Giving up fun now for fun later. That framing made it feel like punishment.</p><p>The truth is simpler. Money I do not spend today gives me options tomorrow. It is the difference between “I have to take this job” and “I can wait for the right one.” Between “I am stuck” and “I have choices.”</p><p>Unspent money is not sitting idle. It is buying freedom.</p><p>That is why I keep both emergency savings and dream savings. Emergency money protects me from disaster. Dream money protects my future possibilities. Both matter.</p><h3>8. This is a season, not a sentence</h3><p>These rules exist to serve me, not shame me.</p><p>I am not living like this forever. This is the student season. The building season. The season where margin is limited.</p><p>One day I will have a higher income, fewer fixed costs, and real breathing room. When that happens, these rules will evolve. My priorities will shift.</p><p>But right now, discipline creates the foundation that freedom is built on.</p><p>This is not restriction.<br> It is protection.<br> And it is temporary.</p><h3>Why this matters</h3><p>Most personal finance advice assumes you have discretionary income. It assumes margin. It assumes that after covering the essentials, there is money left to allocate.</p><p>But if you are a student working part time or a recent graduate managing loans and entry level pay, that is not your reality. Your money is assigned before it even hits your account. Rent, loans, utilities, groceries. It is all spoken for.</p><p>In that context, traditional budgeting advice does not just fail. It convinces you that the problem is your discipline when the actual problem is your math.</p><p>You cannot budget your way out of not having enough. But you can protect what you do have by being ruthlessly clear about what it is for.</p><p>That is what this code does.</p><p>It gives me a framework for operating in a season without margin, without pretending that reality is a moral failure.</p><p>This is my money code for 2026. It will not be yours exactly. But if you are in the same season I am in, maybe it is a starting point.</p><p>The goal is not perfection.<br> It is clarity.</p><p>And clarity, it turns out, is worth more than discipline ever was.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c502d662b57d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Importance of Domain Knowledge in Data Science]]></title>
            <link>https://medium.com/@nicmtisi/the-importance-of-domain-knowledge-in-data-science-a0704a22ed1c?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/a0704a22ed1c</guid>
            <category><![CDATA[domain-knowledge]]></category>
            <category><![CDATA[feature-engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[exploratory-data-analysis]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Fri, 31 Oct 2025 15:28:01 GMT</pubDate>
            <atom:updated>2025-10-31T15:28:01.187Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/636/1*8lETjlp5LE0w-JeqOlNKww.png" /></figure><p>Ever since I got into the field of data analytics and data science, my main goal was to solve problems. I focused on building models, finding p-values, identifying important variables, and applying algorithms. I spent most of my time working on feature engineering, running regressions, and evaluating model performance. To me, that was what data science was all about.</p><p>But I have recently come to a realization that has changed how I see everything. Having domain knowledge about your dataset actually sets you apart. It makes you enjoy the subject more. It helps you relate to the problem you are trying to solve in a deeper way. When you understand the context behind your data, you stop treating analysis as a mechanical process and start seeing it as a way to understand real-world systems.</p><p>Domain knowledge helps you ask better questions. You no longer just look for correlations; you start to think about why those relationships exist. It helps you engineer features that actually represent something meaningful in the real world. It also keeps you from making wrong conclusions because you begin to recognize when the data might be misleading due to how it was collected or structured.</p><p>When you have domain knowledge, you interpret your findings with more clarity. You can explain results in a way that makes sense to people in that field. You also gain more respect for the data itself because you understand what each variable represents and what each missing value or outlier might mean.</p><p>In my recent work with real datasets, I have seen how much difference this makes. When you work with flight data, you understand that delays are not just numbers but real events caused by weather, scheduling, or human error. When you work with health data, you realize that every record represents a person and a condition that affects their life.</p><p>Now, before I start analyzing, I make sure I understand the field the data comes from. I read about it, I study the background, and I try to think like someone in that environment. This approach has made my work more thoughtful and my insights more meaningful.</p><p>As data scientists, we often focus on technical skills, but domain understanding is what turns analysis into insight. Models and metrics are tools, but they only make sense when they’re grounded in real-world context. Having domain knowledge makes you a better problem solver, a clearer communicator, and a more responsible analyst.</p><p>In the end, data science is not just about numbers or code. It’s about understanding the reality those numbers represent. Domain knowledge connects data to purpose, and once you understand that, your entire approach to analytics changes.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a0704a22ed1c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Data to Decisions: Building the Insurance Profitability Explorer]]></title>
            <link>https://medium.com/@nicmtisi/from-data-to-decisions-building-the-insurance-profitability-explorer-ace71889e691?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/ace71889e691</guid>
            <category><![CDATA[insurtech]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[insurance]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-analysis]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Tue, 05 Aug 2025 20:49:11 GMT</pubDate>
            <atom:updated>2025-08-05T20:49:11.005Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Fi3jNopeaT_dW6sIqGLj8w.png" /></figure><p>The insurance industry faces a recurring challenge: achieving the right balance between premium income and claims payouts. While this seems like a straightforward business problem, the reality is often complicated by fragmented data sources, legacy systems, and a lack of cohesive analytical tools. As a result, companies struggle to pinpoint which areas of the business are truly profitable, leading to missed opportunities in underwriting, pricing, and risk management.</p><p>As a data analyst passionate about solving real-world problems, I set out to tackle this gap using data visualization and analytics.</p><p><strong>Identifying the Gap</strong></p><p>Many insurers have the data they need, but it’s often siloed, underutilized, or poorly visualized. This creates blind spots that prevent organizations from making proactive, data-driven decisions. I wanted to build a tool that not only brought all the relevant data together but also helped uncover actionable insights for strategic improvement.</p><p><strong>The Solution: Insurance Profitability Explorer</strong></p><p>To address this challenge, I developed the Insurance Profitability Explorer, an interactive dashboard built in Power BI using a simulated dataset of 5,000 insurance policies.</p><p>The objective was clear: create a visual, intuitive, and interactive tool that allows users to explore financial performance across product types, customer segments, regions, underwriters, and sales channels.</p><p><strong>Key Features and Technical Highlights</strong></p><ul><li><strong>Custom DAX Measures:</strong> I created robust DAX formulas to calculate Total Premium, Total Claims, Total Profit, and Loss Ratio. These measures form the analytical backbone of the dashboard.</li><li><strong>Strategic Layout:</strong> The dashboard begins with high-level KPIs and then allows users to drill into more granular insights, such as profit by underwriter or product type.</li><li><strong>Interactive Slicers:</strong> Users can filter the data by region, product type, underwriter, or sales channel to explore specific performance areas in real time.</li></ul><p><strong>Data Model and Key Columns</strong> The analysis for this project was built upon a single table containing 5,000 policies. The key columns used to build the dashboard and drive the analysis include:</p><ul><li>policy_id: A unique identifier for each policy.</li><li>policy_start_date: Used to track performance over time.</li><li>premium: The amount paid for a policy.</li><li>claims: The amount paid out for a claim.</li><li>commission: The amount paid to the sales channel.</li><li>product_type: The type of insurance (e.g., Auto, Home, Life).</li><li>customer_segment: The customer category (e.g., Individual, Family).</li><li>region: The geographical location of the policy.</li><li>channel: The sales channel (e.g., Online, Agent).</li><li>underwriter: The specific underwriter responsible for the policy.</li></ul><p><strong>Insights and Impact</strong></p><p>The dashboard reveals several key challenges — and more importantly, it points to practical solutions.</p><p><strong>Problem 1: Fragmented Profitability and Risk</strong></p><p>The company appears to have a healthy profit, but the loss ratio sits at 44.9%. A closer look shows that while products like Home and Life are highly profitable, Health insurance operates on much thinner margins.</p><ul><li><strong>Solution:</strong> Reallocate resources to the most profitable product lines or re-evaluate pricing strategies for underperforming products like Health.</li><li><strong>Impact:</strong> A more optimized product portfolio and improved underwriting profitability.</li></ul><p><strong>Problem 2: Underwriter Performance Gaps</strong></p><p>The Loss Ratio by Underwriter visual reveals a wide variation in performance. TrustCare has a significantly higher loss ratio than its peers, while LibertyLife outperforms in both profitability and risk management.</p><ul><li><strong>Solution:</strong> Investigate the underwriting and claims processes at LibertyLife to identify best practices, and apply those insights to underperforming teams like TrustCare.</li><li><strong>Impact:</strong> Greater consistency in underwriting performance and lower overall risk.</li></ul><p><strong>Problem 3: Sales Channel Optimization</strong></p><p>The dashboard shows that Online and Agent channels contribute the most to total profit, while the Partner channel underperforms.</p><ul><li><strong>Solution:</strong> Use the dashboard to analyze the Partner channel by region and product_type, identify where it performs best, and replicate those strategies across other segments.</li><li><strong>Impact:</strong> A more focused and effective sales strategy that aligns channel efforts with profitability goals.</li></ul><p><strong>Built for Real-World Scale</strong> Although this project used simulated data, it was designed to scale in real-world environments. The dashboard structure supports more complex relationships between underwriters, products, and channels, making it adaptable for growing insurance operations.</p><p><strong>Conclusion</strong> The Insurance Profitability Explorer is more than just a dashboard — it’s a strategic decision-making tool. It turns disconnected data into a unified story, helping insurance companies shift from reactive management to proactive strategy.</p><p>By highlighting product performance, underwriting quality, and sales channel efficiency, this solution empowers business leaders to make smarter, data-driven decisions.</p><p><strong>View the Live Demo</strong> <a href="https://app.powerbi.com/groups/me/reports/be19d713-b093-4aa9-8f08-92d9f1ff8001/db7ef0c54b6fbc660132?experience=power-bi">Explore the dashboard here</a></p><p>Have questions or interested in collaborating? Feel free to reach out on <a href="https://www.linkedin.com/in/nicolette-mtisi">LinkedIn</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ace71889e691" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Chronic Care Analytics: Revolutionizing Urban Health with Data]]></title>
            <link>https://medium.com/@nicmtisi/chronic-care-analytics-revolutionizing-urban-health-with-data-d34cb11726d7?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/d34cb11726d7</guid>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Wed, 23 Jul 2025 01:40:14 GMT</pubDate>
            <atom:updated>2025-07-23T01:40:14.648Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R5REQ1rlQ5ksbNVIpnHZPQ.jpeg" /><figcaption>model</figcaption></figure><p>Chronic diseases like diabetes, hypertension, asthma, and heart disease are a growing concern, especially in urban areas, influenced by lifestyle, environmental factors, and healthcare disparities. To effectively address this, healthcare organizations require a robust system for tracking, analyzing, and responding to these trends among urban populations.</p><p>Trio Analytics has developed a comprehensive “Chronic Care Analytics” project designed to provide such a system. This initiative focuses on enabling healthcare providers, researchers, and policymakers to track critical data points, including diagnosis rates, treatment plans, hospital visits, medication usage, and patient outcomes for chronic diseases. The ultimate goal is to facilitate better resource allocation, targeted interventions, and preventive health programs to improve urban population health.</p><h3>The Foundation: Understanding the Data Model</h3><p>Our solution is built upon a well-structured database, with key entities capturing essential information:</p><ul><li><strong>Person:</strong> Stores personal details and medical history.</li><li><strong>Disease:</strong> Details various diseases, including intensity, type, and source.</li><li><strong>Indication:</strong> Represents symptoms related to diseases.</li><li><strong>Medicine:</strong> Catalogs medications used for treatment.</li><li><strong>Location:</strong> Captures geographical data related to patients or diseases.</li><li><strong>Race:</strong> Defines different racial categories of individuals.</li><li><strong>Race Disease Propensity:</strong> Illustrates the relationship between race and disease predisposition.</li><li><strong>Diseased Patient:</strong> An associative entity linking patients with their diagnosed diseases.</li></ul><p>These entities are interconnected through logical relationships, such as a person having multiple diseases, a disease having various indications, and medicines being prescribed based on specific indications. The data model also accounts for the propensity of certain diseases within different racial groups.</p><h3>Building the Engine: OLTP Database and Core Queries</h3><p>We implemented a fully functional Online Transaction Processing (OLTP) database using PostgreSQL. This OLTP database is engineered for daily operational tasks like tracking patient cases, managing facilities, and monitoring medication performance, ensuring efficient querying of transactional data.</p><p>To demonstrate the capabilities of our OLTP system, we ran several key queries:</p><ul><li><strong>Patients and Diagnosed Diseases:</strong> Listing all patients and their diagnosed chronic diseases, along with start dates, end dates, and severity values.</li><li><strong>Medicines and Treated Diseases:</strong> Identifying medicines and the diseases they treat, including their effectiveness percentages.</li><li><strong>Patients per City:</strong> Counting the number of patients residing in each city.</li><li><strong>Diseases with Highest Severity:</strong> Pinpointing diseases with the highest recorded severity values.</li><li><strong>Disease Propensity by Race:</strong> Revealing the propensity of various diseases across different racial categories.</li></ul><p>We also ensured the integrity of the database by demonstrating how new patient information is added and updated, and how integrity constraints prevent invalid data entry, such as adding a person with a non-existent race code.</p><h3>Diving Deeper: The Dimensional Database for Analytics</h3><p>For in-depth analytical capabilities, we created a dimensional database with a grain at the “Person-Diagnosis-Date Level”. This means each record in ourfact_disease_diagnosis table represents a single disease diagnosis for a patient on a specific date, encompassing details like disease type, severity, treatment, and effectiveness.</p><p>This granular level allows for:</p><ul><li><strong>Detailed Person Tracking:</strong> Enabling in-depth tracking of patient histories, treatments, and outcomes.</li><li><strong>Disease Progression Monitoring:</strong> Facilitating the study of disease progression patterns and treatment effectiveness over time.</li></ul><p>Our dimensional model includes several supporting dimension tables:</p><p>dim_date, dim_person, dim_disease, dim_medicine, dim_locationand dim_race. These dimensions provide context without breaking the fact table&#39;s grain.</p><h3>ETL Process: From OLTP to Analytics</h3><p>We utilized an ELT (Extract-Load-Transform) process to populate the dimensional schema from the OLTP database. Data was extracted from the normalized OLTP tables (e.g person, disease, medicine) and loaded into the new warehouse schema, which houses the dimensional model (fact and dimension tables). Transformations, such as matching keys, cleaning formats, handling missing relationships, and calculating aggregates, were then performed within the warehouse using SQL scripts. This shift prepares the system for fast, flexible analytical queries, slicing and dicing across multiple business dimensions, and feeding data to visualization tools.</p><h3>Unlocking Insights: Analytical Queries and Visualizations</h3><p>Our analytical queries on the dimensional model provide crucial insights:</p><ul><li><strong>Medicine Effectiveness by Disease Type:</strong> Analyzing the average effectiveness of medicines based on disease type.</li><li><strong>Race Most Affected by High-Intensity Diseases:</strong> Identifying which racial groups are most impacted by diseases with high intensity levels.</li><li><strong>City with Highest Average Disease Severity:</strong> Pinpointing geographical areas with the highest average disease severity.</li></ul><p>These analytical capabilities are further enhanced by an interactive dashboard built from the dimensional data model. This dashboard visualizes key insights, including:</p><ul><li><strong>Age Distribution:</strong> Showing how different age groups are affected by diseases, aiding in tailored prevention and treatment strategies.</li><li><strong>Disease Risk by Race (Propensity):</strong> Highlighting racial groups with higher diagnosis counts, indicating potential health disparities or genetic susceptibilities.</li><li><strong>Prevalent Diseases:</strong> Identifying the most common diseases within the population to focus healthcare priorities.</li><li><strong>Diagnoses by Location:</strong> Monitoring geographic patterns to pinpoint hotspots and support resource allocation.</li><li><strong>Average Effectiveness by Medicine:</strong> Demonstrating treatment effectiveness across disease types to support evidence-based decisions.</li></ul><p>These visual insights empower healthcare providers and policymakers to make data-driven decisions, target at-risk populations, improve treatment protocols, and optimize funding allocation.</p><p>It’s important to note that all data used in this project is synthetic and was generated solely for demonstration purposes to showcase the system’s analytical capabilities. The values presented are mock data designed to simulate realistic patterns.</p><h3>Future Considerations: NoSQL, Graph Databases, and Cloud Architecture</h3><p>While our current solution leverages a relational database (PostgreSQL), we also considered alternative database structures:</p><ul><li><strong>NoSQL (MongoDB):</strong> Offers flexible schemas, denormalized design, and avoids joins, prioritizing fast reads and scalable storage. This contrasts with relational databases’ focus on consistency and structure.</li><li><strong>Neo4j (Graph Database):</strong> Models data as nodes and relationships, ideal for complex relationships and graph-based queries, such as connecting a person to a disease, a disease to a medicine, or a person to a location.</li></ul><p>For robust and scalable deployment, we envision a Batch + Real-Time Lambda Architecture in AWS. This architecture includes:</p><ul><li><strong>Batch Layer:</strong> Storing raw data in S3, processing with AWS Glue or EMR, and loading into Redshift or RDS Postgres.</li><li><strong>Speed Layer:</strong> Capturing real-time events using Amazon Kinesis and processing updates instantly with AWS Lambda.</li><li><strong>Serving Layer:</strong> Delivering analytics via Amazon QuickSight, with optional Redis Elasticache for fast query responses.</li></ul><p>This AWS setup would incorporate multi-AZ deployments for resilience, IAM permissions for security, and auto-scaling and serverless compute for performance.</p><p>Furthermore, Snowflake offers significant advantages for a growing data-intensive healthcare analytics application due to its separation of compute and storage, fully managed service, high concurrency, native semi-structured data support, and advanced features like time travel and secure data sharing. Snowflake’s capabilities provide greater scalability, ease of management, and advanced analytics compared to a traditional Postgres-based warehouse.</p><p>By leveraging these technologies and a data-driven approach, the Chronic Care Analytics project by Trio Analytics aims to significantly improve urban population health through informed decision-making and targeted interventions.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d34cb11726d7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a BERT-Based Sentiment Classifier for Airline Tweets: My Step-by-Step Journey]]></title>
            <link>https://blog.gopenai.com/building-a-bert-based-sentiment-classifier-for-airline-tweets-my-step-by-step-journey-52bd079c0454?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/52bd079c0454</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[distilbert]]></category>
            <category><![CDATA[sentiment-analysis]]></category>
            <category><![CDATA[hugging-face]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Thu, 10 Jul 2025 06:37:06 GMT</pubDate>
            <atom:updated>2025-07-11T17:53:57.525Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/1*PoepdyD52RCl98UMjseqtw.png" /><figcaption>Hugging face template of my model</figcaption></figure><p>I’ve been diving deep into large language models (LLMs) lately, and to really get hands-on, I decided to revisit sentiment analysis — something I’d previously tackled using classic NLP techniques. If you’ve read my earlier articles, you know I once built a <a href="https://medium.com/@nicmtisi/decoding-user-feedback-my-twitter-sentiment-analysis-project-with-python-da7f2ba8d513">sentiment classifier</a> using traditional NLP pipelines : tokenization, TF-IDF, and a simple classifier like logistic regression.</p><p>This time, I wanted to push further and see what modern transformer models could do. So, I built a sentiment analysis model for airline tweets using BERT, leveraging the Hugging Face Transformers library.</p><h3>Why Sentiment Analysis and Why BERT?</h3><p>Sentiment analysis is a classic NLP task, and airline tweets are a goldmine of real-world, emotionally charged text. I wanted to see how much better a modern transformer-based model like BERT could perform compared to my old school TF-IDF and logistic regression pipelines.</p><h3>1. Data Preparation</h3><p>I started with a CSV file of airline tweets, focusing on the text and airline_sentiment columns. Here’s what I did:</p><ul><li>Loaded the data using pandas.</li><li>Encoded sentiment labels (negative, neutral, positive) into numbers using LabelEncoder—because machine learning models need numbers, not strings.</li><li>Split the data into training and test sets (80/20 split) to make sure I could evaluate performance fairly.</li></ul><pre>import pandas as pd<br>from datasets import Dataset<br>from sklearn.preprocessing import LabelEncoder<br><br>df = pd.read_csv(&#39;Tweets.csv&#39;)[[&#39;text&#39;, &#39;airline_sentiment&#39;]].rename(columns={&#39;airline_sentiment&#39;: &#39;label&#39;})<br>le = LabelEncoder()<br>df[&#39;label&#39;] = le.fit_transform(df[&#39;label&#39;])<br>dataset = Dataset.from_pandas(df)<br>dataset = dataset.train_test_split(test_size=0.2, seed=42)<br></pre><h3>2. Tokenization with BERT</h3><p>Unlike traditional NLP, BERT expects text to be tokenized into subword units and padded/truncated for batch processing. I used the DistilBERT tokenizer for this:</p><pre>from transformers import AutoTokenizer<br><br>tokenizer = AutoTokenizer.from_pretrained(&quot;distilbert-base-uncased&quot;)<br><br>def tokenize_function(example):<br>    return tokenizer(example[&quot;text&quot;], padding=&quot;max_length&quot;, truncation=True)<br><br>tokenized_dataset = dataset.map(tokenize_function, batched=True)</pre><h3>3. Model Setup and Training</h3><p>I initialized a DistilBERT model for sequence classification (three sentiment classes), set up training arguments, and used accuracy as my evaluation metric. Early stopping was enabled so the model wouldn’t overfit.</p><pre>from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer, EarlyStoppingCallback<br>import evaluate<br><br>model = AutoModelForSequenceClassification.from_pretrained(&quot;distilbert-base-uncased&quot;, num_labels=3)<br>accuracy = evaluate.load(&quot;accuracy&quot;)<br><br>def compute_metrics(eval_pred):<br>    logits, labels = eval_pred<br>    predictions = logits.argmax(axis=-1)<br>    return accuracy.compute(predictions=predictions, references=labels)<br><br>training_args = TrainingArguments(<br>    output_dir=&quot;./bert-sentiment&quot;,<br>    evaluation_strategy=&quot;epoch&quot;,<br>    save_strategy=&quot;epoch&quot;,<br>    load_best_model_at_end=True,<br>    metric_for_best_model=&quot;eval_loss&quot;,<br>    greater_is_better=False,<br>    per_device_train_batch_size=16,<br>    per_device_eval_batch_size=16,<br>    num_train_epochs=10,<br>    weight_decay=0.01,<br>    logging_dir=&quot;./logs&quot;,<br>    logging_steps=10,<br>)<br><br>trainer = Trainer(<br>    model=model,<br>    args=training_args,<br>    train_dataset=tokenized_dataset[&quot;train&quot;],<br>    eval_dataset=tokenized_dataset[&quot;test&quot;],<br>    tokenizer=tokenizer,<br>    compute_metrics=compute_metrics,<br>    callbacks=[EarlyStoppingCallback(early_stopping_patience=1)],<br>)<br></pre><p>Training the model was surprisingly smooth, and I watched the validation accuracy improve with each epoch.</p><h3>4. Evaluation</h3><p>After training, I evaluated the model on the test set. Here’s a sample of the results I got:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PVp4yOTIapf1UxQHD30sQg.png" /><figcaption>Results</figcaption></figure><p>I was happy to see accuracy consistently above 83% — a big step up from my previous classic NLP models</p><h3>5. Saving and Inference</h3><p>I saved the best-performing model and tokenizer, then wrote a function to predict the sentiment of any new tweet. This approach is much more robust than my old pipeline.</p><pre>from transformers import AutoModelForSequenceClassification, AutoTokenizer<br>import torch<br><br>model = AutoModelForSequenceClassification.from_pretrained(&quot;bert_sentiment_final&quot;)<br>tokenizer = AutoTokenizer.from_pretrained(&quot;bert_sentiment_final&quot;)<br><br>def predict_sentiment(text):<br>    inputs = tokenizer(text, return_tensors=&quot;pt&quot;, padding=True, truncation=True, max_length=128)<br>    inputs = {k: v.to(model.device) for k, v in inputs.items()}<br>    model.eval()<br>    with torch.no_grad():<br>        outputs = model(**inputs)<br>        predicted_class = outputs.logits.argmax(dim=1).item()<br>    return le.inverse_transform([predicted_class])[0]<br><br># Example:<br>print(predict_sentiment(&quot;I love the friendly service on this airline!&quot;))  # Output: positive</pre><h3>6. Deployment to Hugging Face Hub</h3><p>To make the model accessible to others, I uploaded it to the Hugging Face Hub. This involved logging in with my API token, creating a repository, and pushing both the model and tokenizer.</p><pre>from huggingface_hub import create_repo<br><br>repo_name = &quot;bert-sentiment-nic&quot;<br>create_repo(repo_name, private=False)<br>model.push_to_hub(&quot;Nicolettem/bert-sentiment-nic&quot;)<br>tokenizer.push_to_hub(&quot;Nicolettem/bert-sentiment-nic&quot;)</pre><h3>7. Building a Gradio Web App</h3><p>Finally, I wanted to let anyone interact with my model, so I built a simple Gradio web app. Users can enter any airline-related sentence and instantly get a sentiment prediction.</p><pre>import gradio as gr<br><br>label_map = {0: &quot;negative&quot;, 1: &quot;neutral&quot;, 2: &quot;positive&quot;}<br><br>def predict_sentiment(text):<br>    inputs = tokenizer(text, return_tensors=&quot;pt&quot;, truncation=True, padding=True, max_length=128)<br>    with torch.no_grad():<br>        outputs = model(**inputs)<br>        logits = outputs.logits<br>        predicted_class_id = int(torch.argmax(logits, dim=-1))<br>    return label_map.get(predicted_class_id, &quot;Unknown&quot;)<br><br>demo = gr.Interface(<br>    fn=predict_sentiment,<br>    inputs=gr.Textbox(lines=3, placeholder=&quot;Enter airline feedback here...&quot;),<br>    outputs=&quot;text&quot;,<br>    title=&quot;Airline Sentiment Classifier&quot;,<br>    description=&quot;Enter a sentence about an airline, and this model predicts the sentiment: positive, neutral, or negative.&quot;<br>)<br><br>demo.launch()</pre><h3>What I Learned</h3><p>Switching from classic NLP to LLMs like BERT was a game-changer. The workflow was smoother, the results were better, and deployment was easier than ever. If you’re curious about the code or want to try it yourself, you can follow my process as a template for your own sentiment analysis projects.</p><p>l have also provided the links to my work below</p><ul><li><a href="https://huggingface.co/Nicolettem/bert-sentiment-nic">Nicolettem/bert-sentiment-nic · Hugging Face</a></li><li><a href="https://huggingface.co/spaces/Nicolettem/bert_sentiment_demo">Bert Sentiment Demo - a Hugging Face Space by Nicolettem</a></li></ul><p>l share more about my journey on <a href="https://www.linkedin.com/in/nicolette-mtisi">LinkedIn </a>, happy to connect!</p><p>Let me know if you have questions or want to see more!</p><p><em>Thanks for reading! If you found this helpful, feel free to share or leave a comment below.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52bd079c0454" width="1" height="1" alt=""><hr><p><a href="https://blog.gopenai.com/building-a-bert-based-sentiment-classifier-for-airline-tweets-my-step-by-step-journey-52bd079c0454">Building a BERT-Based Sentiment Classifier for Airline Tweets: My Step-by-Step Journey</a> was originally published in <a href="https://blog.gopenai.com/">GoPenAI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Zim Roots, American Rhythms: My Funny, Tough, and Transformative Culture Shock]]></title>
            <link>https://medium.com/@nicmtisi/from-sekuru-sam-to-sparkling-water-my-zimbabwean-take-on-american-culture-shock-b912c5e352ec?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/b912c5e352ec</guid>
            <category><![CDATA[american-culture]]></category>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[new-york]]></category>
            <category><![CDATA[zimbabwe]]></category>
            <category><![CDATA[culture]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Wed, 02 Jul 2025 19:35:22 GMT</pubDate>
            <atom:updated>2025-07-02T19:44:37.498Z</atom:updated>
            <content:encoded><![CDATA[<p>A love letter to the awkward, hilarious, and humbling moments of navigating a new culture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uUrc7o-e66jwJ7kEzuyBaQ.jpeg" /><figcaption>From the vibrant markets of Zimbabwe to the sprinting streets of New York, my culture shock journey began.</figcaption></figure><blockquote>“Call me Sam”</blockquote><p>That’s what he said. A man clearly in his 60s, wearing a golf cap and orthopedic shoes. Back home in Zimbabwe, we would’ve called him Mr. Sam or, at the very least, Sekuru Sam. I mean the man has lived through decades, probably raised a whole village of children, and here he is, telling me to just call him Sam.<br>I blinked.<br>That was my first taste of culture shock in America. No titles, no formalities, just vibes.</p><p><strong>The Casualness Was Confusing</strong><br>In Zimbabwe, respect isn’t optional. It’s stitched into how we speak, greet, and hand over things. You don’t just say “hi” to an elder like you’re ordering a sandwich. You clear your throat, throw in a “good morning,” and maybe even bow slightly with your voice.<br>But in America? Everyone’s on a first-name basis. Even your professor might say, “Call me Dave.” Dave? Excuse me?<br>And don’t get me started on the time someone called their mom by her first name. I had to sit down and pray.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/413/1*YiK9BG_J9vczVC6H5R-LnQ.png" /><figcaption>When your professor says, “Call me Dave,” but your Zimbabwean instincts scream, “Sekuru Dave!”</figcaption></figure><p><strong>The Sparkling Water Betrayal</strong><br>One time, I went to an Italian restaurant, and the waiter asked, “Still or sparkling?”<br>I thought sparkling sounded fancy, like water that glows in candlelight or something from a rom-com.<br>No one warned me.</p><blockquote>“Why does it feel like the water is fighting me?”</blockquote><p>That first sip was spiritual. Not in a good way. Why is it fizzy? Why does it feel like the water is fighting me?<br>Back home, water is humble. Cold, still, and loyal. It doesn’t come with surprises.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/681/1*1aAFzrTL2opjXoM2q9-0Ug.png" /><figcaption>Sparkling water: the fizzy betrayal l never saw coming.</figcaption></figure><p><strong>Ranch, Chipotle, and Other Culinary Confusion</strong><br>Ranch dressing. Americans treat ranch like it’s a love language.<br>• Pizza? Ranch.<br>• Wings? Ranch.<br>• Fries? Ranch.<br>I tried it once and had to double-check if it was expired. It’s giving seasoned yogurt with identity issues.<br>Then there’s Chipotle.<br>At first, I thought it was a spice. Then I found out it’s a restaurant. Then I found out it’s still kind of a spice. Or maybe a flavor? Honestly, I still don’t know.<br>All I know is I walked into a Chipotle, someone asked, “Black beans or pinto?” and I just said “Yes” because I had no idea what that meant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/462/1*v18xpOgh3DTa81tpESZEjA.png" /><figcaption>Chipotle or spice? America’s food scene remains a mystery.</figcaption></figure><p><strong>The McDonald’s Moment I Will Never Forget</strong><br>We walked into McDonald’s, ready to order, and that’s when it happened — the communication breakdown of the century.<br>The guy behind the counter said something. I didn’t hear him. I asked him to repeat. He did. Still nothing. I responded. He looked confused.</p><p>I even considered asking for a pen and paper so we could write things down like a hostage negotiation.<br>After a few painful rounds of “Sorry?” and “What was that?” we gave up and left. Quietly. Our hunger was replaced by secondhand embarrassment and emotional fatigue.<br>That silence on the way out? Heavy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/492/1*FahEKwyh4Hz4DVl1sHjYmQ.png" /><figcaption>When ordering at McDonald’s feels like a galactic language barrier</figcaption></figure><p><strong>The Accent Thing Is Real</strong><br>Sometimes, my accent turns a simple conversation into a marathon. I went to a networking event in New York, one of those crowded meetups I sometimes attend, hoping to connect with new people. I started chatting with someone, but every sentence was a struggle. I’d say something, and their face would freeze, like they were buffering. I’d repeat myself — two, three times — louder, slower, clearer, until they finally nodded. It’s exhausting, honestly. You start doubting your English, even though you’ve spoken it your whole life. You flatten your voice, strip away the Shona rhythm, just to be understood. It’s not just a language barrier; it’s a piece of your identity you have to pause to get by.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/542/1*wm1ZfUVaR16B3UB9f5OBeQ.png" /><figcaption>When your accent stops the conversation, and you start questioning your English</figcaption></figure><p><strong>New York Moves Like It Has No Pause Button</strong><br>Then came the day I visited Manhattan.<br>New York doesn’t walk — it sprints.<br>Everyone is power walking like they’re late for a meeting with Beyoncé. No eye contact. No lingering. No slowing down to appreciate the sun.<br>I stood still for two seconds to check Google Maps, and at least five people zoomed past me like I was scenery.<br>Back home, we greet people. We walk with purpose and peace. In New York, if you don’t walk with urgency, you’ll get swallowed by the sidewalk.<br>It’s not that people are rude, they’re just in their own world, locked into their mission. Focused. Efficient.<br>But for someone used to rhythm and flow, it’s a full-on sprint culture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/552/1*D62nkI9q8l7_aJ224_eq8A.png" /><figcaption>New York sidewalks: where standing still makes you scenery</figcaption></figure><p><strong>And Yet, I’m Growing</strong><br>Even with all the awkward encounters, weird food, fizzing water, and accidental sidewalk collisions, I’ve grown. I’ve laughed. I’ve learned.</p><blockquote>“Culture shock isn’t just about being confused. It’s about being humbled.”</blockquote><p>It’s about seeing your own culture through new eyes. And it’s about learning to hold space for both the new and the familiar.<br>Some days, I miss home deeply. I miss hearing Shona in the background. I miss the rhythm of Zim life, slower, warmer, known.<br>But I also like the person I’m becoming here. A little braver. A little louder. Still me, just… evolving.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sjmTqWgY7wzbGSKaqeQO1A.jpeg" /></figure><p><strong>So, Here I Am</strong><br>• No, I still don’t know what Chipotle is.<br>• Yes, I still flinch at sparkling water.<br>• And yes, I’ll forever call elders “Sir” or “Ma’am,” even if they tell me not to.<br>But I’ll keep showing up. Accent, confusion, Google Maps, and all.</p><p>Have you ever experienced culture shock? What was your “sparkling water” moment? Drop a comment below and let’s laugh (or cry) together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b912c5e352ec" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decoding User Feedback: My Twitter Sentiment Analysis Project with Python]]></title>
            <link>https://medium.com/@nicmtisi/decoding-user-feedback-my-twitter-sentiment-analysis-project-with-python-da7f2ba8d513?source=rss-2195b4c1bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/da7f2ba8d513</guid>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[sentiment-analysis]]></category>
            <category><![CDATA[python-programming]]></category>
            <category><![CDATA[nlp]]></category>
            <dc:creator><![CDATA[nicolette mtisi]]></dc:creator>
            <pubDate>Fri, 27 Jun 2025 01:36:16 GMT</pubDate>
            <atom:updated>2025-06-27T01:36:16.215Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>Ever wondered what people really think about airlines on Twitter? As a scientist-in-training passionate about data-driven UX analytics, I built a Python-based NLP project to classify airline tweets as positive, neutral, or negative with ~80% accuracy. Inspired by my entrepreneurial ventures, I wanted to explore how data can uncover user sentiments to improve experiences. In this article, I’ll walk you through my process, from preprocessing tweets to visualizing insights, and share how this project fuels my goal to create impactful, user-centric solutions.</p><h3>Project Overview</h3><p>This project uses the Twitter Airline Sentiment dataset from Kaggle (~14,000 tweets) to classify sentiments using Logistic Regression. Built with Python, NLTK, scikit-learn, and Matplotlib/Seaborn, it achieves ~80% accuracy and visualizes user feedback trends. The goal? Extract insights to enhance UX analytics and understand customer experiences.</p><h3>Technical Approach</h3><h3>Data Loading and Exploration</h3><p>I loaded the dataset with pandas and explored its structure, finding a majority of negative sentiments — hinting at customer pain points.</p><h3>Text Preprocessing</h3><p>Tweets are messy! I used NLTK to:</p><ul><li>Clean text (remove URLs, mentions, hashtags with regex).</li><li>Tokenize, remove stopwords, and lemmatize words.</li><li>Convert text to features with TF-IDF vectorization.</li></ul><p>Here’s the preprocessing code:</p><pre>def preprocess_text(text):<br>    text = text.lower()<br>    text = re.sub(r&#39;http\S+|@\w+|#\w+|[^a-zA-Z\s]&#39;, &#39;&#39;, text)<br>    tokens = word_tokenize(text)<br>    stop_words = set(stopwords.words(&#39;english&#39;))<br>    tokens = [word for word in tokens if word not in stop_words]<br>    lemmatizer = WordNetLemmatizer()<br>    return &#39; &#39;.join([lemmatizer.lemmatize(word) for word in tokens])</pre><h3>Model Training</h3><p>I split the data (80% train, 20% test) and trained a Logistic Regression model with scikit-learn, achieving ~80% accuracy.</p><h3>Visualization</h3><p>Using Matplotlib and Seaborn, I created:</p><ul><li>A bar plot showing sentiment distribution.</li><li>A confusion matrix to evaluate model performance.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/863/1*HH0GKTZiAdKZ1rG17pohyQ.png" /><figcaption>Bar Plot</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/779/1*-ktX_gd-VKrFkt5AXM-SDw.png" /><figcaption>confusion matrix</figcaption></figure><h3>Results and Insights</h3><p>The model achieved ~80% accuracy, with strong performance on negative sentiments (reflecting the dataset’s skew). The bar plot revealed most tweets were negative, suggesting areas for airline UX improvements. Challenges like noisy data were overcome with robust preprocessing. These insights are valuable for UX analytics, highlighting user frustrations to inform design.</p><h3>Future Improvements</h3><p>Next steps include:</p><ul><li>Testing advanced models like BERT for better accuracy.</li><li>Integrating the Twitter API for real-time analysis.</li><li>Building Tableau dashboards for interactive visualizations.</li></ul><h3>Conclusion</h3><p>This project was a deep dive into NLP and user feedback analysis, blending my technical skills with my entrepreneurial drive to solve real-world problems. I’m excited to apply these insights to UX analytics and beyond. Check out the code on GitHub:</p><p>[https://github.com/nic-stack/Twitter-Sentiment-Analysis],</p><p>connect with me on [<a href="https://www.linkedin.com/in/nicolette-mtisi">https://www.linkedin.com/in/nicolette-mtisi</a>],</p><p>or share your thoughts below!</p><p><em>By Nicolette, a data scientist passionate about UX analytics and social impact.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da7f2ba8d513" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>