<?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 Siddharthan P S on Medium]]></title>
        <description><![CDATA[Stories by Siddharthan P S on Medium]]></description>
        <link>https://medium.com/@siddharthanps.1?source=rss-569d9e9fc2a3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*k7ws4KHuZtv6kZyS6qGWSA.jpeg</url>
            <title>Stories by Siddharthan P S on Medium</title>
            <link>https://medium.com/@siddharthanps.1?source=rss-569d9e9fc2a3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 16:27:48 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@siddharthanps.1/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[Predicting Customer Lifetime Value (LTV) with Python]]></title>
            <link>https://medium.com/@siddharthanps.1/predicting-customer-lifetime-value-ltv-with-python-379e2e15b6fd?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/379e2e15b6fd</guid>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Thu, 04 Sep 2025 22:31:25 GMT</pubDate>
            <atom:updated>2025-09-04T22:51:58.296Z</atom:updated>
            <content:encoded><![CDATA[<p>I got my hands on a digital wallet dataset recently, not the usual textbook fare, but something that feels real. Think customers, transactions, income levels, all the stuff that tells a story. I wanted to see: who among these users are the real MVPs in terms of lifetime value (LTV)? And can a handful of Python tricks in Colab reveal that?</p><p>Here’s what I did, how I cleaned it up, and why it felt more like detective work than coding.</p><h4>When Data Speaks, You Just Listen</h4><p>At first glance, the numbers looked as chaotic as rush-hour traffic in Bengaluru. Skewed distributions. Missing values. A few customers spending way more than anyone else. Instead of diving straight into models, I decided to <em>see</em> the data, not just read it.</p><p>A histogram of LTV opened my eyes, it wasn’t a neat bell curve but a long winding trail with a few big spenders out there alone at the tail.</p><p>A heatmap of correlations whispered what I already suspected: Income, Total Spent, Average Transaction Value, these weren’t random factors; they were playing lead roles.</p><p>And then a scatter plot, with points coloured by income level, suddenly patterns emerged. Clusters of high spend, groups of low-income, low-transactions. That plot alone made me pause and smile: wow, there’s something real here.</p><h4>Cleaning the Noise Felt Like Clearing the Fog</h4><p>I knew I couldn’t ignore those outliers, they distort patterns like streetlights in a fog. So I used a simple trifecta: Z-score filtering, IQR method, and eyeballing the weird ones. Point was to bring the data closer to what felt reasonable.</p><p>Once filtered, the heatmap looked cleaner, correlations made more sense. A bit like tuning a radio until that station finally sounds crisp.</p><h4>A Few Smart Tweaks, Not Reinventing the Wheel</h4><p>Instead of sticking to raw things like “Total Spent”, I created a new feature: <strong>Average Spend per Transaction</strong>. Felt intuitive: someone might have low income, but if their average transaction is high, they’re still valuable.</p><p>Then I standardised everything so features played nicely together. That meant models wouldn’t unfairly favour variables just because of their scale.</p><h4>Modelling Without the Stress</h4><p>I tried a range: a simple Linear Regression (solid baseline), Support Vector Regression (handles nuance better), Gradient Boosting (a reliable workhorse), and even a neural network for the curious bit.</p><p>At the end of the day, the neural network edged ahead, but not by a cosmic margin. Gradient boosting held its ground, and I liked that, it felt dependable, lighter, and still strong.</p><h4>What Stood Out</h4><p><strong>Cleaning matters more than coding.</strong> If the data is messy, even the fanciest model falters. <strong>Feature engineering beats feature piling.</strong> That little average spend feature punched way above its weight. <strong>Compare models, don’t just pick.</strong> One dataset, many contenders. Let the numbers decide. <strong>Neural nets are cool, but boosting is rock solid.</strong> Don’t always chase novelty. Sometimes, tried and tested works better.</p><h4>If I Were to Journal This, Photos I’d Include</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TXfzVu6Z4NFhQBmkn6H92g.png" /><figcaption>A correlation heatmap (pre cleaning)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r6mFcJprY4ZB71SvTQDudw.png" /><figcaption>A correlation heatmap (post cleaning)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XR1_tkm3Z1NiggjNLE9L9g.png" /><figcaption>Distribution of LTV (original vs. filtered)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Vfd8QruOSmous8I981HnfQ.png" /><figcaption>How close did the model get? Almost hugging the ideal line</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sX2aNnkDj8hECiaMJzCkPQ.png" /><figcaption>Seeing the misses: where the model struggled versus where it nailed it</figcaption></figure><p>Each visual should carry a little caption: “Here’s why I paused”, “This told me to clean up”, “See that cluster? That was a signal”, “Here’s the impact of a new feature”, “Models don’t all perform the same”.</p><h4>To Wrap Up Something More Than Code</h4><p>What really struck me was this: working on LTV isn’t just about predictions. It’s about shaping business decisions, how much to spend acquiring users, who to keep happy. The real win wasn’t in chasing a few decimal points more in accuracy, it was about understanding the data’s voice and letting that guide the model.</p><p>So if you’re working with customer data, or just curious about ML in fintech, you don’t need a fancy setup. Start with Colab, clean wisely, engineer a few intuitive features, and let the models earn their place. Your next step? Let’s get that dashboard live, or share this with the team that sets marketing budgets.</p><p><strong>Models I used to summarize and complete it :</strong></p><ul><li><strong>Linear Regression</strong><br> The simplest baseline. It tried to fit a straight line relationship between inputs (like Total Spent, Avg Transaction Value, Income Level) and LTV. Not very flexible, but a good starting point.</li><li><strong>Support Vector Regression (SVR)</strong><br> This can capture non-linear relationships by using kernels. Works well when the data isn’t linearly separable, but it can be slower on bigger datasets.</li><li><strong>Gradient Boosting Regressor</strong><br> An ensemble of decision trees built sequentially, where each tree corrects the errors of the previous one. This gave you strong performance without needing deep learning.</li><li><strong>Neural Network (TensorFlow/Keras)</strong><br> A multi-layer model built with Keras. By stacking dense layers with activations and regularisation, it was able to capture more complex patterns in spending and transactions. This model achieved the highest R-squared in your notebook.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=379e2e15b6fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What ETF Price Trends Taught Me About Investing — Through the Eyes!]]></title>
            <link>https://medium.com/@siddharthanps.1/what-etf-price-trends-taught-me-about-investing-through-the-eyes-086960d019d1?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/086960d019d1</guid>
            <category><![CDATA[etf]]></category>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[etf-investing]]></category>
            <category><![CDATA[data-visualization]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Sun, 18 May 2025 21:38:42 GMT</pubDate>
            <atom:updated>2025-05-19T16:32:38.754Z</atom:updated>
            <content:encoded><![CDATA[<h3>What ETF Price Trends Taught Me About Investing — Through the Eyes of a Data Enthusiast</h3><blockquote>“Markets are never wrong, only people are.”<em> — Jesse Livermore</em></blockquote><blockquote><em>tAs someone who juggles between code and coffee breaks, I recently stumbled upon a massive dataset that had the scent of opportunity historical ETF prices. Now, I’m no stock guru or YouTube finance influencer shouting “Buy the dip!” But when a dataset has over </em><strong><em>3.8 million rows</em></strong><em>, my inner data geek angle just </em>had<em> to investigate.</em></blockquote><h3>🧳 The Dataset: Bigger Than My Grocery List</h3><p>We’re talking about Exchange-Traded Funds (ETFs), which are like mutual funds, but with the cool factor of stocks. Tradable, transparent, and often a good bet for passive investing.</p><p>The CSV file I got had these columns:</p><ul><li>fund_symbol — unique codes like AAA, XYZ, etc.</li><li>price_date — daily historical records</li><li>open, high, low, close, adj_close — your usual suspects</li><li>volume — the number of units traded that day</li></ul><pre>df = pd.read_csv(&#39;ETF prices.csv&#39;)<br>print(&quot;Dataset shape:&quot;, df.shape)</pre><p>The sheer size of it? <strong>3.8 million rows × 8 columns.</strong> Enough to keep me busy through an entire IPL match.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5XkvOMflfUQ72knh4-_aSw.png" /></figure><h3>👀 First Impressions Matter</h3><p>So naturally, I started with the basics — df.describe(), df.info(), and value counts on the fund_symbol. The idea was simple: before I ask the data to predict tomorrow, I better understand its yesterday.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F59DNvRrKWUxvLYEV3Flrg.png" /></figure><p>Turns out, some funds had <strong>thousands of daily entries</strong>, while others just popped in for a few dates. That already tells you something — some ETFs are consistent performers, others maybe experimental or discontinued.</p><h3>🎨 Visualizing Trends: When in Doubt, Plot it Out</h3><p>Let’s be honest — no one enjoys staring at numbers for hours. So I pulled out matplotlib and seaborn, and started plotting time-series for popular funds. The daily closing prices showed tiny fluctuations, but over a span of months? You could clearly see upward or downward momentum.</p><p>This part was exciting — it was like watching a thriller unfold, but with charts.</p><h3>🧠 Prediction Wali Baat: Enter Machine Learning</h3><p>Now for the fun part. I wanted to see if we can use ML to predict something like whether tomorrow’s closing price would go up or down.</p><p>So I cooked up a mini pipeline:</p><pre>from sklearn.linear_model import Ridge, Lasso, LinearRegression<br>from sklearn.ensemble import RandomForestClassifier<br>from sklearn.model_selection import train_test_split<br>from sklearn.preprocessing import StandardScaler</pre><p>After cleaning the data and creating lag features (like yesterday’s price, moving averages, etc.), I split it into training and testing sets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n-6hH2AvvyuXc23WVqsuTA.png" /></figure><p>One surprise? Even a simple <strong>Linear Regression</strong> model gave halfway decent results. Not enough to quit your day job and become a trader, but definitely better than flipping a coin.</p><h3>💡 Learnings from the Trenches</h3><p>Let me be real building models on stock market data isn’t about achieving 100% accuracy. It’s about <strong>finding signals in the noise</strong>.</p><p>Here’s what stood out:</p><ul><li><strong>Volume matters</strong> — ETFs with higher volume had more predictable trends.</li><li><strong>Recent performance &gt; old data</strong> — Including too much history sometimes made models worse.</li><li><strong>Feature engineering is king</strong> — Raw prices won’t do much; it’s how you shape them that counts.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*faKv5GA6EW_2OXNqpnafRg.png" /></figure><h3>🛣️ What’s Next?</h3><p>In Part 2 (yes, this is just the trailer), I’ll show:</p><ul><li>How I built a full-fledged ML pipeline to predict price direction</li><li>Visual insights from Random Forest importance plots</li><li>Whether any of it is actually <strong>profitable</strong> in the real world</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*arCNTb_ZrGoAyJw2u6malQ.png" /></figure><p>Because let’s be honest no one cares about RMSE if it doesn’t help your demat account 😅</p><h3>☕ Final Thoughts Over a coffee</h3><p>If you’ve made it this far, you’re clearly curious and that’s the best thing you can be in today’s noisy world. Whether you’re from a commerce background or a curious coder like me, the market has lessons for everyone.</p><p>Until then <strong>Keep coding, keep questioning, and never trust a trend without a time axis.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=086960d019d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[‍♂️ Detecting Money Laundering Using Python — My Hands-On Attempt]]></title>
            <link>https://medium.com/@siddharthanps.1/%EF%B8%8F-detecting-money-laundering-using-python-my-hands-on-attempt-5e772aef6e8c?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e772aef6e8c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[anti-money-laundering]]></category>
            <category><![CDATA[money-laundering-case]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Tue, 13 May 2025 19:51:53 GMT</pubDate>
            <atom:updated>2025-05-21T23:52:56.454Z</atom:updated>
            <content:encoded><![CDATA[<h3>🕵🏽‍♂️ Detecting Money Laundering Using Python &amp; ML — My Hands-On Attempt</h3><p>Lately, I’ve been diving into datasets that feel a little more “real-world.” Not just Titanic or Iris-type classics, I wanted something that has practical use in areas like finance, fraud, or compliance.</p><p>That’s when I came across a synthetic <strong>Anti-Money Laundering (AML)</strong> dataset on <a href="https://www.kaggle.com/datasets/berkanoztas/synthetic-transaction-monitoring-dataset-aml">Kaggle</a>. It’s designed to mimic financial transactions with a mix of normal and suspicious activity. No real names, no real banks but the patterns are modeled on actual cases. That was more than enough for me to jump in.</p><h3>Cleaning the Junk First</h3><p>The first thing I saw was — the data was messy. Typical of any raw transaction log. Amounts were all over the place, and the distribution was completely skewed. So I had to log-transform a few features to bring things under control.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4wTjZzZUEwd-YX9eh8yV3A.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JzwlCDL02-luZWvOpb18bA.png" /><figcaption>Image shows the class distribution</figcaption></figure><p>That alone made a huge difference. I could finally see some actual shape to the data instead of just spikes.</p><h3>What’s Going On in These Transactions?</h3><p>Next step understanding the behaviour. I plotted the daily volume of transactions, just to see if there’s any seasonality or anomaly. And yeah, some months did stand out.</p><p>But the real kicker was when I visualised <strong>suspicious transaction amounts </strong>huge spikes, randomly scattered, but definitely not normal.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-6BI_Xj3pBQFDeKNHjGgIA.png" /><figcaption>Avg Number of transactions</figcaption></figure><p>Some accounts were also getting flagged repeatedly way more than the average. Could be fake accounts or mule accounts. Whatever the case, it stuck out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LQ3mI_K8fKSkRdcoCvB8fg.png" /><figcaption>Suspicious Transaction Amounts</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PuNHccAfTlHWTFhWtwzPjg.png" /></figure><h3>What Type of Laundering Are They Doing?</h3><p>Laundering isn’t random. There are methods to the madness called “typologies.” So I visualised which ones were being used most.</p><p>Among normal transactions, you have usual things like cash deposits and small fan-outs. But in the suspicious ones, <strong>structuring</strong>, <strong>smurfing</strong>, and <strong>layered transactions</strong> dominated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*W8vHw5lJtC59TysMV0A-YA.png" /></figure><p>This part gave me a lot of insight into how frauds are structured not just catching them, but understanding the playbook.</p><h3>How Laundering Varies by Payment Type</h3><p>One thing I was really curious about was the <strong>channel</strong> people use to launder money. I mean, are they doing it through cash, cards, transfers, or something else?</p><p>So I broke it down by payment method — and this was eye-opening.</p><p>As expected, regular things like debit cards, cheques, and ACH transfers had high volumes overall. But when I compared <strong>laundering counts</strong>, one type stood out like a sore thumb <strong>cross-border transactions</strong>.</p><p>Despite not being the most frequently used method, it had the <strong>highest laundering count by far</strong>. That’s a huge red flag. If this was real-world data, compliance teams would be all over this.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/740/1*PL0VWM88H9EmK1_AWJKegg.png" /></figure><p>Then I looked at how these payment methods trended over time — and again, cross-border was consistently high. But interestingly, other methods like <strong>cash deposits</strong> and <strong>withdrawals</strong> also had small spikes in some months.</p><p>This kind of monthly breakdown really helped me see <strong>which methods picked up suspicious activity over time</strong>, and which ones stayed relatively stable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dmly5qICVSK8wjmeVIqrjw.png" /></figure><p>If I were working at a bank, this is the kind of chart I’d have on my dashboard daily. One glance and you know where to look.</p><h3>Time to Test the Model</h3><p>With all this understanding, I trained a model. Nothing fancy. Just to see whether a machine learning model could flag suspicious transactions correctly based on the patterns.</p><p>Results? Pretty solid.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*BahmEmFpGQF9z62Av6hzyg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/834/1*50XnXlrhpuNdfrFLOURzCg.png" /></figure><p>The ROC curve showed an AUC of <strong>0.81</strong> — not bad at all. And when I tuned the threshold to aim for 90% true positive rate, the <strong>confusion matrix</strong> showed the trade-off: some false positives, but still caught the frauds.</p><h3>What I Learned</h3><p>This project wasn’t about just getting good accuracy. It taught me how <strong>real-world problems need real understanding.</strong></p><ul><li>Data cleaning matters more than anything.</li><li>Suspicious behaviour has <strong>visual patterns</strong>. Plot first, model later.</li><li>Typologies and country flow help you understand what’s actually going on.</li><li>Even with a basic model, if your features are solid, results will speak.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e772aef6e8c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ I Built a Credit Card Fraud Detection Dashboard Using Big Data-Here’s What Happened]]></title>
            <link>https://medium.com/@siddharthanps.1/i-built-a-credit-card-fraud-detection-dashboard-using-big-data-heres-what-happened-3639514b5415?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/3639514b5415</guid>
            <category><![CDATA[creditcardfrauddetection]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Sun, 04 May 2025 22:43:16 GMT</pubDate>
            <atom:updated>2025-05-05T01:06:44.532Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the past few weeks, I’ve been building something I’ve always wanted to — a proper <strong>fraud detection dashboard</strong> that runs on real data, predicts fraud, and looks like a SaaS product. It’s not a tutorial project. It’s a full product.</p><p>And now it’s live:<br> 🔗 <a href="https://creditcardfraud.info/">https://creditcardfraud.info</a></p><p>In this post, I’ll walk you through the stack I used, how I handled big data, the problems I faced, and some takeaways. If you’re someone learning ML or backend/frontend dev, this is for you.</p><h3>💡 What I Built</h3><p>The idea was simple: upload a CSV of credit card transactions, run predictions using an ML model, and display the results on a clean dashboard.</p><p>But once you scale this up to 1 million+ records and aim for real-world usage… the simplicity goes away very quickly.</p><h3>🧠 The ML Model That Did the Job</h3><p>I trained a <strong>Random Forest Classifier</strong> using an anonymized real-world dataset of credit card transactions.</p><p>Why Random Forest?</p><ul><li>It handles imbalanced datasets quite well (which fraud detection almost always is).</li><li>It doesn’t overfit as easily as other classifiers.</li><li>It gives <strong>probabilistic outputs</strong>, which I used to rank transaction risk.</li></ul><p>But what made it even better? I added <strong>undersampling techniques</strong> (like NearMiss) to rebalance the classes, and did proper <strong>feature selection</strong> to remove noisy variables. That uplifted the precision-recall curve significantly.</p><p>In plain terms: the model stopped blindly calling every transaction “non-fraud” just because it was rare.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/691/1*k3x3lJuelz2AppI8XoBccw.png" /></figure><p>We started catching real anomalies — with over <strong>96% precision</strong> on frauds during testing.</p><p>After Uplift:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/790/1*JMOqoyZG2Lu_mH8Mxcxr9Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/348/1*96k7N1iv9xODEKyyI_uQOg.png" /></figure><p>The model was saved using joblib as fraud_model.pkl, and the backend loads it in-memory for fast response.</p><h3>⚙️ Stack Overview</h3><p><strong>Frontend:</strong></p><ul><li>React + TypeScript</li><li>Tailwind CSS + Shadcn UI</li><li>Framer Motion for animations</li></ul><p><strong>Backend:</strong></p><ul><li>Flask + Pandas</li><li>REST API to receive CSV and return fraud predictions</li><li>ML model with scikit-learn</li></ul><p><strong>Hosting:</strong></p><ul><li>Render (Flask API)</li><li>IONOS (React frontend)</li><li>Namecheap (Domain)</li></ul><h3>😤 Real Problems I Faced</h3><h3>1. Huge Files, Small RAM</h3><p>My 1M-row CSV crashed everything at first. Flask couldn’t handle it natively. I used an cloud application platform to support Flask Backend.<br> I fixed it with:</p><ul><li>stream=True for file upload</li><li>Chunked reading via pandas</li><li>Lightweight inference logic</li></ul><h3>2. CORS Issues</h3><p>Classic issue — frontend and backend on different domains.<br>Spent a full day debugging preflight OPTIONS issues until I got the headers right.</p><h3>3. Model-Feature Mismatch</h3><p>My model was trained on 30 features, but user data sometimes had 28 or 29. So I added model_features.json to force column alignment.</p><p>📊 Dashboard Features</p><ul><li>Upload a CSV with transaction data</li><li>Get a full table with fraud probability</li><li>Highlighted frauds, sorted by risk</li><li>Interactive charts showing fraud trend</li><li>Paginated and responsive layout (mobile friendly too!)</li></ul><p>You don’t need Python to use it. That was the point.</p><h3>🧪 What I Learned</h3><ul><li><strong>ML is 30% of the work</strong> — the rest is infra, UI, and usability.</li><li><strong>Test with real data sizes</strong> from Day 1.</li><li><strong>Don’t ignore UX</strong> — even fraud predictions look better in charts.</li><li><strong>Deployment is a whole skillset</strong> — not just git push.</li></ul><h3>✅ Try It Out</h3><p>Want to test it?</p><p>👉 <a href="https://creditcardfraud.info/">https://creditcardfraud.info</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uPCoc11k9gmCISf6drWr5w.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2CUuaVqTyfwSW-tB1KbsKA.png" /></figure><p>Upload your own data (or use sample ones), and see the predictions live — in seconds. No setup. No code. Just fraud scores.</p><h3>🙌 Wrapping Up</h3><p>This was the most satisfying project I’ve worked on.<br>Not because of the tech — but because it’s real, usable, and deployed.</p><p>If you’re learning ML or full-stack dev, build something like this. One real project teaches more than 10 Coursera certificates.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3639514b5415" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[CI/CD for Cloud ML Models: Automating Model Deployment with Azure Pipelines]]></title>
            <link>https://medium.com/@siddharthanps.1/ci-cd-for-cloud-ml-models-automating-model-deployment-with-azure-pipelines-ecb6a8c42b7d?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/ecb6a8c42b7d</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Sat, 03 May 2025 19:40:14 GMT</pubDate>
            <atom:updated>2025-05-03T19:40:14.233Z</atom:updated>
            <content:encoded><![CDATA[<p>When I first started working on machine learning projects, deployment always felt like an afterthought. You spend weeks tuning your model, get decent metrics, and then suddenly you’re asked: “Can you put this into production?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*alOt0szca6O6vfsd" /></figure><p>The first few times, I did what most of us do manual deployment. It worked… until it didn’t. The process was fragile, error-prone, and time-consuming. That’s when I realised I needed to treat ML deployment like software engineering — with automation, version control, and reliability. That’s where <strong>CI/CD</strong> (Continuous Integration and Continuous Deployment) came in.</p><h3>Why CI/CD for ML Models?</h3><p>Unlike traditional software, ML systems have more moving parts: the model, the code, the data, and the infrastructure. Any of these can change independently. CI/CD helps tie all of this together in a repeatable, testable workflow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*IwAVBdLcfwt1e6Pe" /><figcaption>A flowchart comparing traditional software CI/CD with ML CI/CD, showing additional components like data, models, and experiments.</figcaption></figure><p>Instead of manually retraining and deploying a model every time something changes, I now rely on pipelines that:</p><ul><li>Trigger automatically when code or data updates</li><li>Run checks and validations</li><li>Deploy the latest stable version to production</li><li>Roll back if anything breaks</li></ul><p>This reduces manual effort, avoids silly mistakes, and lets me focus on what matters — improving the model.</p><h3>Why Azure Pipelines?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*H5BXUCoCoAkP3gtT.png" /><figcaption>Azure DevOps Pipeline dashboard</figcaption></figure><p>There are plenty of tools out there — GitHub Actions, Jenkins, GitLab CI — but I leaned into <strong>Azure Pipelines</strong> for a few reasons:</p><ul><li>Smooth integration with Azure Machine Learning and Container Registry</li><li>Clean UI for managing pipeline steps and logs</li><li>Flexibility for both code-first and GUI-based users</li><li>Easy secrets management and access control</li></ul><p>It took me a few iterations to get comfortable, but once I set it up, the pipeline became part of my daily workflow.</p><h3>The Workflow That Works for Me</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*D3q5H1htxZ-DBUIU" /><figcaption>A clean pipeline diagram showing a DevOps cycle</figcaption></figure><p>My typical CI/CD flow for ML looks something like this:</p><ol><li>Push code to the repository</li><li>The pipeline gets triggered automatically</li><li>It runs checks, like whether the model meets accuracy thresholds or if data validation passes</li><li>If everything’s good, the model is registered and deployed as a versioned API</li><li>If something breaks, I get alerts — and nothing goes live</li></ol><p>Once this is set up, updates become a matter of just pushing changes. The rest happens like clockwork.</p><h3>Real-World Lessons</h3><p>Through trial and error, here’s what I’ve learned:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*z1IqoEPwh-XYCOeP" /></figure><ul><li><strong>Don’t skip testing.</strong> Even small data changes can lead to big prediction errors.</li><li><strong>Logs matter.</strong> A good logging setup saves hours of debugging.</li><li><strong>Keep everything modular.</strong> Pipelines are easier to manage when each step is independent.</li><li><strong>Monitor after deployment.</strong> Just because the model deployed successfully doesn’t mean it’s working as expected.</li></ul><p>Most importantly, <strong>don’t wait until the end to think about deployment</strong>. Build with deployment in mind from Day 1.</p><h3>Final Thoughts</h3><p>CI/CD for ML isn’t just a buzzword — it’s a mindset shift. It turns ML from a research prototype into a reliable product. If you’re just starting out, it might feel overwhelming, but trust me, it pays off. Your models will thank you later.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*blLqMFI4asF7OoCp" /></figure><p>I’ll probably write a follow-up post soon on lessons from managing model versions and handling live updates. But if you’ve been relying on manual deployment, consider this your sign to automate it.</p><p>It’s one of the most useful things I’ve ever done for my workflow — and I wish I’d started earlier.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ecb6a8c42b7d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We, the Robots? What I Learnt from Studying Human-Robot Teaming as an Engineer]]></title>
            <link>https://medium.com/@siddharthanps.1/we-the-robots-what-i-learnt-from-studying-human-robot-teaming-as-an-engineer-cf145029bc98?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf145029bc98</guid>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Fri, 02 May 2025 22:02:54 GMT</pubDate>
            <atom:updated>2025-05-02T22:02:54.040Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>“We’re not just building robots. We’re designing how humans will work with them.”</blockquote><p>When I started this project, I wasn’t thinking about writing a blog post or impressing anyone. I just wanted to explore a question that has been nagging me for a while: <em>What is the actual role of humans in an automated future?</em></p><p>Like most engineering students, I had heard about robots in manufacturing, defense, and surgery. But what I hadn’t thought about deeply was this robots aren’t just tools anymore. We’re starting to treat them like teammates. That’s where things get interesting… and tricky.</p><h3>Why Human-Robot Teaming Caught My Curiosity</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/720/0*G0iMZDw3uEpnRvBa" /></figure><p>To be honest, I have no real-world industry experience. But I’ve always been curious about technology <em>with a human angle</em>. That’s why I picked this topic for my Systems Engineering course project at NYU.</p><p>I wasn’t interested in just how robots work. I wanted to understand how humans and robots <strong>work together</strong> — especially when things go wrong.</p><p>In every field, from automobile assembly to telemedicine to bomb disposal, robots are entering human workspaces. But are we designing these systems properly? Or are we just rushing machines into human zones without thinking through the consequences?</p><h3>Human-Robot Teaming: It’s Not What You Think</h3><p>It’s not just about a robot picking up a box or stitching a patient. Real teaming means:</p><ul><li><strong>Sharing environments</strong></li><li><strong>Communicating intent</strong></li><li><strong>Dividing tasks smartly</strong></li><li><strong>Handling unexpected events together</strong></li></ul><p>And most importantly — <strong>trust</strong>.</p><p>That’s the hardest part. How do you design trust into a machine?</p><h3>Systems Engineering Changed My Thinking</h3><p>Before this course, I used to think Systems Engineering was just glorified project management. But this course flipped that idea.</p><p>I learnt it’s actually a way of thinking. A structure to make sense of complex, messy problems. It forces you to look at:</p><ul><li>Who are the stakeholders?</li><li>What are the needs?</li><li>Where can things break?</li><li>What trade-offs are we making?</li></ul><p>In human-robot teaming, this SE mindset helped me tie together everything from safety systems and feedback loops to ethical dilemmas and technical risks.</p><h3>I Studied the Good, the Bad, and the Avoidable</h3><p>Instead of building a robot, I did what a student could do best <strong>study cases</strong>, dig into research, and <strong>map everything</strong> to the SE lifecycle.</p><p>Here are a few that stood out:</p><h3>✅ BMW Cobots — Success Story</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eC7UdL-pYOJf-YYM" /><figcaption>Source: LinkedIn — Universal Robots</figcaption></figure><p>In BMW’s factories, cobots help with heavy lifting and repetitive tasks. But what impressed me was how well the system was <strong>integrated</strong>.</p><ul><li>Real-time safety sensors</li><li>Visual indicators to signal intent</li><li>Comfort-focused task design using Human Factors Engineering</li></ul><p>They didn’t just throw robots on the floor they <em>engineered</em> the teamwork.</p><h3>✅ Da Vinci Surgical Robot — Precision with Control</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*j9vEZD840xg3KkFt" /><figcaption>Source: docquity</figcaption></figure><p>This robot doesn’t work alone. It amplifies a surgeon’s hand through robotic precision. What I found fascinating:</p><ul><li>Tremor-cancellation</li><li>Instant freeze on power loss</li><li>Better visuals than the human eye</li></ul><p>This is what trust through design looks like.</p><h3>❌ Warehouse AMRs — Failure to Adapt</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/780/0*-n6khGtv8xgmbitl" /><figcaption>Source: designworldonline.com</figcaption></figure><p>Autonomous Mobile Robots in warehouses sounded futuristic. But many failed in real-world tests.</p><ul><li>They couldn’t handle unpredictable human movement</li><li>Navigation systems broke down in cluttered spaces</li><li>Feedback to humans was almost non-existent</li></ul><p>No system-level thinking = high failure rate.</p><h3>❌ Military Robots — Failure Under Pressure</h3><p>In one case, a bomb-disposal robot just stopped mid-operation due to signal jamming. There was <strong>no fallback mode</strong>. In another, latency in teleoperation caused near-miss situations.</p><p>A soldier’s life was at stake, but the system wasn’t built with enough redundancy or autonomy. That’s not a tech problem. That’s a <strong>systems problem</strong>.</p><h3>The Real Technical Struggles No One Talks About</h3><p>As I dug deeper, I noticed something: most failures weren’t about bad robots. They were about bad systems design.</p><ul><li><strong>Sensor fusion issues</strong> — Getting data is easy. Trusting it is hard.</li><li><strong>Weak communication layers</strong> — No point in a message that arrives late.</li><li><strong>Poor real-world testing</strong> — Robots that pass lab tests crumble in chaos.</li></ul><p>We’re obsessed with AI and mechatronics but sometimes forget the <em>boring but essential</em> parts of integration.</p><h3>Ethics Wasn’t Just a Chapter. It Was Everywhere.</h3><p>I didn’t expect to care about ethics this much. But this project forced me to think beyond code and control loops.</p><ul><li>Who is responsible when a robot fails?</li><li>How do you design <em>fear</em> out of a system?</li><li>What if the robot has dataset bias?</li></ul><p>I read a case where workers avoided walking near cobots out of fear even when they were proven safe. That’s not just a social issue. That’s a <strong>design flaw</strong>.</p><h3>My Takeaway: Don’t Just Build Robots. Build Relationships.</h3><p>Through this project, I didn’t become a robotics expert. But I did become someone who sees robots not as gadgets, but as <strong>participants in a system</strong>.</p><p>We can’t afford to build impressive robots that fail in real life. We need to design systems where <strong>humans feel safe, respected, and empowered</strong>.</p><p>As engineers, we must go beyond innovation. We must design for reality and for people.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf145029bc98" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cloud Migration Strategies for Systems-A Real-World Playbook]]></title>
            <link>https://medium.com/@siddharthanps.1/cloud-migration-strategies-for-systems-a-real-world-playbook-d48d5c3b0190?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d48d5c3b0190</guid>
            <category><![CDATA[cloud-migration-strategy]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Thu, 01 May 2025 21:40:38 GMT</pubDate>
            <atom:updated>2025-05-01T21:40:38.340Z</atom:updated>
            <content:encoded><![CDATA[<p>Cloud migration isn’t just about lifting and shifting applications from legacy infrastructure to a cloud provider like AWS, Azure, or GCP. It’s a business-critical transformation impacting performance, scalability, security, cost efficiency, and sometimes, even the organisational mindset.</p><p>Having worked on cloud solutions at an MNC and explored real-world data pipelines during my graduate research, I’ve seen that a successful migration strategy goes beyond theory. It’s about context. What works for a startup may not suit a enterprise with monolithic systems. Let’s dive into the nuanced approaches to cloud migration, along with practical tips that I’ve learned the hard way.</p><h3>Understanding the Need for Strategy</h3><p>No two systems are alike. Some are I/O intensive, some are compute-bound, and others carry legacy baggage with tightly coupled architectures. That’s why cloud migration strategies need to be tailored. A good strategy mitigates risks such as:</p><ul><li><strong>Downtime during migration</strong></li><li><strong>Data loss or leakage</strong></li><li><strong>Cost overruns</strong></li><li><strong>Poor post-migration performance</strong></li></ul><h3>The 6 Rs of Cloud Migration: Revisited</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rEgG3h7dHE7EXd35Raazng.png" /></figure><p>You’ve likely come across the 6 Rs of migration, but let me reframe them based on practical relevance:</p><h3>1. Rehost (“Lift and Shift”)</h3><p>Best for: Quick migrations with minimal code changes.</p><p>I once helped move a client’s analytics workload from on-prem to Azure VMs using Docker and Kubernetes. It was fast, but post-migration, we realised that without re-architecting, costs creep up fast.</p><p><strong>Tip:</strong> Use this when time is short, but revisit cost and performance tuning post-migration.</p><h3>2. Replatform (“Lift, Tinker, and Shift”)</h3><p>Best for: Systems that need minor optimisations without full re-architecture.</p><p>At MNC, we containerised an internal API system and hosted it on Azure App Services, swapping SQL Server with Azure Database. Minor changes, massive impact on scalability.</p><p><strong>Watch out for:</strong> Dependency mismatches and database porting nuances.</p><h3>3. <strong>Repurchase</strong></h3><p>Best for: Moving from custom to SaaS (e.g., moving from self-hosted Salesforce to Salesforce Cloud).</p><p>For example, many educational institutions I supported at NYU shifted from legacy CRM tools to Salesforce Cloud. It was smoother in terms of UX, but custom workflows needed redesigning.</p><h3>4. Refactor / Re-architect</h3><p>Best for: Long-term gains with microservices, serverless, or cloud-native transformation.</p><p>This is high-effort but high-reward. Think: decomposing a monolithic app into services with Kubernetes, CI/CD, and managed databases. If you’re a cloud-native enthusiast, this is your playground.</p><p><strong>Pitfall:</strong> Avoid it if your team lacks cloud maturity or DevOps readiness.</p><h3>5. Retain</h3><p>Best for: Systems not ready for cloud or not worth migrating.</p><p>Some systems are better left on-prem (e.g., sensitive data governed by strict compliance). You can always migrate later when regulations loosen or business priorities shift.</p><h3>6. <strong>Retire</strong></h3><p>Best for: Redundant or unused systems.</p><p>Always perform a system audit. You’d be surprised how many cron jobs or services are no longer needed but consume resources.</p><h3>Migration Blueprint: From Planning to Execution</h3><ol><li><strong>Discovery</strong> — Audit all systems, dependencies, network flow.</li><li><strong>Assessment</strong> — Evaluate cost, effort, security impact of each system.</li><li><strong>Pilot</strong> — Start small (e.g., migrate a batch job or internal tool).</li><li><strong>Migrate in Waves</strong> — Prioritise based on criticality and complexity.</li><li><strong>Monitor and Optimise</strong> — Post-migration cloudwatching is key. Use observability tools like Azure Monitor or CloudWatch.</li></ol><h3>Common Pitfalls &amp; Lessons</h3><ul><li><strong>Underestimating network latency:</strong> Especially when hybrid architecture is in place.</li><li><strong>Ignoring hidden costs:</strong> Egress, premium tier services, and idle VMs can shock your billing dashboard.</li><li><strong>Lack of rollback planning:</strong> Always have a rollback or snapshot plan before big moves.</li></ul><h3>Final Thoughts</h3><p>Cloud migration is as much about strategy as it is about technology. It’s not a sprint it’s a system-wide evolution. The best migrations I’ve worked on weren’t the fastest or fanciest ones they were thoughtful, phased, and well-communicated. If you’re starting out, begin with systems that are cloud-ready. Build internal champions. Make DevOps and cost visibility part of the culture.</p><p>Remember, the goal isn’t just “cloud first.” It’s <strong>value first</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d48d5c3b0190" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Visualization Best Practices: Beyond the Basics]]></title>
            <link>https://medium.com/@siddharthanps.1/data-visualization-best-practices-beyond-the-basics-af1beec3926f?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/af1beec3926f</guid>
            <category><![CDATA[data-visualization]]></category>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Thu, 01 May 2025 21:28:46 GMT</pubDate>
            <atom:updated>2025-05-01T21:28:46.161Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Taking your data stories to the next level with advanced visualisation techniques.</em></p><h3>Introduction: Why Visualisation Still Matters</h3><p>We’ve all heard it before <em>a picture is worth a thousand words</em>. But in the data world, a well-crafted chart can be worth a million. With dashboards and KPIs floating around in every company’s Slack channel, we’ve entered an era where just putting a bar chart isn’t enough. We need to go beyond the basics both in design and in intent.</p><p>As someone who’s spent late nights perfecting dashboards and waking up to stakeholder feedback like “Can we make this pop?”, I’ve learnt that data storytelling isn’t just about clean charts it’s about <em>effective communication</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kcFMCBxijGxNlUvJFkwmNw.png" /></figure><h3>1. Tell a Story, Not Just Show a Chart</h3><p>One of the biggest mistakes I made early on was thinking a visual alone would speak for itself. It won’t. You have to <em>guide</em> your audience.</p><p>✔️ Use annotations to highlight insights.<br> ✔️ Always provide context — what changed and why it matters.<br> ✔️ Add a headline that gives away the punchline. Not “Sales Q1 2024” but “North Zone Sales Up 32% Highest in 2 Years.”</p><p>📌 <strong>Tip:</strong> Every chart should answer a business question. If it doesn’t, don’t add it.</p><h3>2. Colours Can Make or Break It</h3><p>Indian readers especially prefer visual clarity many of us still read newspapers in print. So don’t confuse people with rainbow pie charts or ten shades of blue.</p><p>✅ Use colour <em>intentionally</em>. Red for loss, green for growth, it’s psychological.<br> ❌ Don’t overdo it. Three to four colours max, unless you’re showing a heatmap or gradient.<br> 🎨 Use tools like ColorBrewer or material palettes to keep things consistent.</p><h3>3. Choose the Right Chart (Always!)</h3><p>This one seems obvious but is often overlooked.</p><ul><li><strong>Bar/Column Charts</strong>: Good for comparing categories</li><li><strong>Line Charts</strong>: Great for trends over time</li><li><strong>Scatter Plots</strong>: Correlation and distribution</li><li><strong>Heatmaps</strong>: Identifying intensity or clustering</li><li><strong>Sankey/Chord</strong>: For flows and relationships especially in finance or logistics</li></ul><p>🚫 Don’t use pie charts unless you <em>have</em> to and never more than five slices.</p><h3>4. Minimise Chart Junk</h3><p>Design isn’t about adding things it’s about <em>removing the unnecessary</em>.</p><p>Strip off:</p><ul><li>Gridlines (or lighten them)</li><li>Thick borders</li><li>3D effects</li><li>Background patterns</li><li>Cliparts or emojis on dashboards (unless it’s part of the UX on purpose)</li></ul><p>✅ Use white space like a pro. Give your data room to breathe.</p><h3>5. Don’t Let Tools Drive the Story</h3><p>Just because Power BI offers 20 visuals doesn’t mean you need to use all of them.</p><p>Tools should <em>enable</em>, not dictate your design. Always start with a pen and paper. Sketch the insight first, then open Tableau, Looker, or Python.</p><p>I’ve often found better insights using a pandas pivot table than an overstyled dashboard. <em>Less tech, more thinking.</em></p><h3>6. Dashboarding vs Reporting: Know the Difference</h3><p>A dashboard is a <em>monitoring tool</em>. It gives a quick overview, often in real-time. A report is an <em>analysis tool</em>. It dives deep into what happened, why, and what next. Mixing the two will leave your stakeholders confused.</p><p>🎯 Keep dashboards light and consistent. Keep reports detailed and insightful.</p><h3>7. Consider Accessibility (It’s Often Ignored)</h3><p>We don’t talk about this enough. Accessibility isn’t just for screen readers — it’s for <em>everyone</em>.</p><ul><li>Use high-contrast colour schemes</li><li>Add alt text to images</li><li>Avoid relying on colour alone use labels or patterns</li><li>Keep font sizes readable (nothing below 12pt for key text)</li></ul><h3>8. Keep Iterating — Design is a Journey</h3><p>The best visualisations I’ve made were <em>version 3 or 4</em>, not the first.</p><p>Feedback is gold:</p><ul><li>Ask colleagues outside your domain: “Do you get what this means?”</li><li>Run A/B tests on layout, especially for public dashboards or product metrics.</li><li>Sleep on it seriously. You’ll see problems the next morning you missed the night before.</li></ul><h3>Conclusion: Craft, Don’t Just Code</h3><p>In the end, data visualisation is both art and science. You need the maths to calculate the metric, the tools to show it but you need empathy to <em>deliver it right</em>.</p><p>Good charts don’t just show data. They <em>drive action</em>. Great charts do something even better they make you <em>think</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af1beec3926f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM (Large Language Models) in Enterprise AI]]></title>
            <link>https://medium.com/@siddharthanps.1/making-sense-of-large-language-models-in-enterprise-ai-af680bea8fec?source=rss-569d9e9fc2a3------2</link>
            <guid isPermaLink="false">https://medium.com/p/af680bea8fec</guid>
            <dc:creator><![CDATA[Siddharthan P S]]></dc:creator>
            <pubDate>Thu, 01 May 2025 21:05:02 GMT</pubDate>
            <atom:updated>2025-05-01T21:18:08.986Z</atom:updated>
            <content:encoded><![CDATA[<p>Over the last year or so, it’s become impossible to ignore the buzz around Artificial Intelligence, especially with terms like ChatGPT, LLMs, and GenAI becoming household names at least in the tech world. I’m someone who’s been casually following the AI wave, but recently I’ve started digging deeper, especially into how large language models (LLMs) are being used in enterprises.</p><p>So in this blog, I just wanted to share some of my thoughts and understanding on how LLMs are making a real impact in businesses, not just in fancy research labs or as tools for students to write assignments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MY6wZ0v8kK3MTVEtY9U-cQ.png" /></figure><h3>First, what exactly is a Large Language Model?</h3><p>In very simple terms, a large language model (LLM) is a type of AI that has been trained on a huge amount of text data. It learns patterns in language, sentence structure, facts, reasoning, and even creativity to some extent. Once trained, it can answer questions, summarise documents, translate languages, write code, and do a lot more.</p><p>Think of it like a super advanced auto-complete, but not just for your sentences, but for full ideas, conversations, and tasks.</p><h3>Now, how is this useful in an enterprise?</h3><p>This is the interesting part. While it’s easy to get caught up in the fun side of AI (like writing poems or generating memes), businesses are using LLMs in very practical ways.</p><p><em>Here are a few examples I came across:</em></p><h3>1. Customer Support Automation</h3><p>Instead of having hundreds of agents answering the same type of queries, companies are now using LLMs to power chatbots and virtual agents. These bots can handle routine queries instantly, with proper tone and context, and escalate to a human only when needed. It saves time, money, and improves customer satisfaction.</p><h3>2. Document Understanding</h3><p>Enterprises deal with thousands of documents contracts, invoices, legal forms, etc. Earlier, humans had to go through these manually. Now, LLMs can be trained to understand these documents, extract relevant data, summarise them, or even highlight risks and errors. This is a game-changer for sectors like insurance, banking, and legal.</p><h3>3. Knowledge Management</h3><p>A lot of companies have internal knowledge scattered across files, emails, and platforms. LLMs can act as internal search engines or knowledge assistants. So an employee can ask: “How do I apply for medical leave?” and instantly get the right policy from the right document without wasting time.</p><h3>4. Content Generation</h3><p>Marketing teams are using LLMs to draft emails, blogs, product descriptions, and more. Of course, the final edit is still done by humans (as it should be), but the AI can handle the boring first draft part.</p><h3>Challenges? Of course, there are</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BK9o0Fud825TL_fh4gee1A.png" /></figure><p>It’s not all smooth. Enterprises still face issues like:</p><ul><li><strong>Data privacy and security</strong> : Especially with sensitive documents, companies are hesitant to trust third-party models.</li><li><strong>Accuracy and hallucination : </strong>LLMs sometimes give wrong answers very confidently, which can be dangerous in business.</li><li><strong>Cost of implementation</strong> : Not all companies can afford fine-tuning models or integrating them properly into their systems.</li></ul><p>But still, the direction is clear : LLMs are not just a tech experiment anymore. They’re becoming part of the real toolkit for digital transformation.</p><h3>My personal takeaway</h3><p>Before I started reading seriously, I thought LLMs were just fancy toys for coders or something that only startups in Silicon Valley would use. But now I realise that they are practical tools that are slowly getting embedded into the backbone of how modern businesses operate.</p><p>For any student, engineer, or even non-tech professional in India, it’s a good time to at least understand the basics of how this works. You don’t need to be an AI expert, but being aware of how these tools are changing the workplace will definitely give you an edge.</p><p>Thanks for reading!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=af680bea8fec" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>