<?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 Ragesh R on Medium]]></title>
        <description><![CDATA[Stories by Ragesh R on Medium]]></description>
        <link>https://medium.com/@rageshramesh4?source=rss-57ba6bf448e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*12tCuZ1N7_wv-Xdi1j2XzA.jpeg</url>
            <title>Stories by Ragesh R on Medium</title>
            <link>https://medium.com/@rageshramesh4?source=rss-57ba6bf448e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 14:18:06 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@rageshramesh4/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[How self experimentation can help you become a better developer]]></title>
            <link>https://medium.com/@rageshramesh4/how-self-experimentation-can-help-you-become-a-better-developer-4665b34af72f?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/4665b34af72f</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[self-experimentation]]></category>
            <category><![CDATA[experiment]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 02:27:21 GMT</pubDate>
            <atom:updated>2025-12-02T02:27:21.882Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DlSCXBxtoCfa6DWkcxm-tw.png" /></figure><p>After 15 years of building APIs, fixing bugs at 3 AM, and pretending to understand legacy code written by “mysterious former employees,” I thought I had my learning style figured out.</p><p>Then I read Barbara Oakley’s <em>A Mind for Numbers</em>, and it hit me:</p><blockquote>“The best learners are self-experimenters.”</blockquote><p>Not geniuses.</p><p>Not people with photographic memory.</p><p>Just curious humans who tinker — with their own habits.</p><p>And honestly… that changed how I learn, how I code, and how I approach problems.</p><p>Here are the experiments that turned me from “I’ll figure it out eventually” to</p><p>“I can debug this without sacrificing my sanity.”</p><h3>1. The Pomodoro Experiment — My Brain’s Favorite Magic Trick</h3><p>I used to code in long, heroic 5-hour stretches. I thought I was being productive. Turns out I was just tired.</p><p>Barbara talks about focused and diffuse thinking:</p><p><strong>Focused mode</strong> → deep work, solving, coding</p><p><strong>Diffuse mode</strong> → subconscious connections, creativity, breakthroughs</p><p>Pomodoro creates a natural switch between the two.</p><p>So I tried:</p><p><strong>45 minutes of coding → 5 minutes of doing nothing</strong></p><p>(yes, literally nothing)</p><p>At first, it felt like stopping a Netflix show at the climax.</p><p>Painful.</p><p>But the results?</p><ul><li>Fewer stupid bugs</li><li>More “aha moments”</li><li>Less brain-melt at the end of the day</li><li>And shockingly: faster delivery</li></ul><p><strong>Self-Experiment Result:</strong></p><p>Breaks are fuel, not interruptions.</p><blockquote>The pomodoro is actually 25 min but i felt that it was too quick of a time for me to get into the flow so after experimentations i fixed on the 45 min mark. Try it out yourself and see what works for you but to keep it withing an hour.</blockquote><h3>2. The “Explain It to a Rubber Duck” Experiment</h3><p>We all know rubber-duck debugging.</p><p>I tried it for learning new AI concepts.</p><p>I’d “explain” transformers to a <strong>Goku Bobble Head</strong> on my desk like I was teaching a mildly confused pigeon.</p><p>And it worked.</p><p>If I couldn’t explain it simply, I didn’t really understand it.</p><p><strong>Metaphor from Barbara:</strong></p><p>Teaching is how you catch the gaps in your “chunks.”</p><p><strong>Self-Experiment Result:</strong></p><p>Talking out loud makes you smarter (even if your coworkers question your sanity).</p><blockquote>This experiment normally is done with a rubber duck but i like dragon ball z so explaining to my goku bobble head made it more fun for me. Pick any toy that you like to make this more personal.</blockquote><h3>3. The “Night Before Review” Experiment</h3><p>This one came straight out of Oakley’s strategies.</p><p>Instead of cramming 4 hours into one long coding tutorial binge, I’d:</p><ul><li>learn for 30 minutes</li><li>sleep</li><li>review for 10 minutes the next morning</li></ul><p>That tiny next-day review cemented the concept more than the original lesson.</p><p>Turns out sleep is not just for survival — it’s where your brain quietly makes mental “indexes.”</p><p><strong>Self-Experiment Result:</strong></p><p>Future me remembers things better when past me stops being a hero.</p><blockquote>Play around with the learn times and review duration. Different duration works for different people. It wouldnt be a proper experiment if the same duration worked for everyone.</blockquote><h3>4. The “Messy First Draft” Coding Experiment</h3><p>I used to force myself to write beautiful code on the first try.</p><p>Spoiler: the code was neither beautiful nor on time.</p><p>Inspired by the idea of “deliberately bad drafts,” I tried writing:</p><ul><li>ugly code</li><li>with comments like “FIX THIS LATER YOU BUFFOON”</li><li>THEN refactoring</li></ul><p>The freedom of being messy first let me focus on understanding the problem, not impressing my imaginary code reviewer.</p><p>And ironically?</p><p>Cleaner final code.</p><p><strong>Self-Experiment Result:</strong></p><p>Bad first drafts = good final results.</p><blockquote>Messy Code gets you started. You start worrying about the functionality instead of how beautiful your code looks. This is freedom. You can always refactor later.</blockquote><h3>5. The “One Concept a Day” Experiment</h3><p>Oakley emphasizes the power of spaced repetition and chunking.</p><p>So instead of trying to understand <em>all of AI</em> in one weekend (yes, I tried that), I did something radical:</p><p><strong>One concept a day.</strong></p><p>That’s it.</p><p>Example:</p><p>Day 1 → embeddings</p><p>Day 2 → attention</p><p>Day 3 → loss functions</p><p>Day 4 → vector databases</p><p>Day 5 → RAG</p><p>After a month, I felt like Neo seeing the Matrix.</p><p><strong>Self-Experiment Result:</strong></p><p>Tiny daily doses beat massive weekend marathons.</p><blockquote>Dont worry if you skip a day. I have skipped many the key is to get back as fast as possible instead of starting over. If the topic is too big for a day split it into multiple days. Movement over perfection.</blockquote><h3>6. The “Celebrate Small Wins” Experiment</h3><p>This one’s simple but powerful.</p><p>Every time I solved a bug, I’d write it down in a “Wins” file:</p><ul><li>“Fixed WebSocket disconnection bug.”</li><li>“Reverse-engineered a weird regex from 2016.”</li><li>“Finally understood how CDC and Domain driven works.”</li></ul><p>Looking at that list on bad days felt like reading my own personal highlight reel.</p><p><strong>Self-Experiment Result:</strong></p><p>Confidence isn’t a feeling — it’s data.</p><blockquote>Every time i had a small win i would treat myself to comfort food instead of when i felt like having it. This helped me cut down on sugar/junk as well as kept me looking forward to the next win. Reward yourself on what excites you.</blockquote><h3>The Real Lesson: Be Your Own Lab</h3><p>The biggest insight from <em>A Mind for Numbers</em> isn’t a study technique.</p><p>It’s the mindset:</p><blockquote>Treat learning like an experiment — not a performance.</blockquote><p>Try new routines.</p><p>Try different learning styles.</p><p>Try weird things.</p><p>Some will fail hilariously.</p><p>Some will work beautifully.</p><p>But each experiment makes you understand yourself a little better.</p><p>And that, more than any design pattern or AI model, is what turns a good developer into a great one.</p><h3>Final Thought</h3><p>After 15 years of building software, here’s my honest conclusion:</p><blockquote><strong><em>Self-experimentation is the ultimate developer superpower. </em></strong><em>It turns your brain into a system you can debug, optimize, and upgrade.</em></blockquote><p>And the best part?</p><p>No production deploys are harmed in the process. 😄</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4665b34af72f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Buzzwords That Finally Made Sense]]></title>
            <link>https://medium.com/@rageshramesh4/ai-buzzwords-that-finally-made-sense-ca447277b030?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/ca447277b030</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agentic-rag]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Tue, 25 Nov 2025 19:12:53 GMT</pubDate>
            <atom:updated>2025-11-25T19:12:53.035Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TQjAcPfb4PS_KdECJAY_GQ.png" /></figure><p>Fifteen years into coding, I thought I had a pretty good handle on tech jargon.</p><p>Then AI exploded. Suddenly every conversation sounded like a Silicon Valley spellbook:</p><blockquote><em>We’re using an agentic AI backed by an LRM with RAG on top of a vector DB, orchestrated via MCP and future-ready for MoE scaling.</em></blockquote><p>I’d nod like I understood it…</p><p>But inside my brain, it looked like that meme of a cat staring at algebra equations.</p><p>So I decided to decode these terms in a way my brain actually likes: <strong>metaphors</strong>.</p><p>And honestly? It changed everything.</p><p>Here are the AI terms I wish someone had explained to me this way earlier.</p><h3>1. Agentic AI — The Self-Driven Intern Who Actually Gets Things Done</h3><p>Traditional AI feels like a calculator: you ask a question, it answers.</p><p>Agentic AI is different.</p><p>It’s like hiring an intern who <em>doesn’t wait for instructions</em>.</p><p>You say:</p><blockquote><em>“Help me plan a trip to Japan.”</em></blockquote><p>An Agentic AI would:</p><ul><li>research flights</li><li>compare prices</li><li>create an itinerary</li><li>check weather</li><li>remind you to renew your passport</li></ul><p>…because it knows the <strong>goal</strong> and can take autonomous steps to get there.</p><p><strong>Metaphor:</strong></p><p>Agentic AI is like giving your computer a <em>to-do list</em>, not a <em>question</em>.</p><h3>2. LRM (Large Reasoning Model) — The Sherlock Holmes of AI Models</h3><p>We’re all familiar with LLMs (Large Language Models) — they predict the next word.</p><p>LRMs are built to <strong>reason</strong>, not just autocomplete.</p><p>If LLMs are parrots, LRMs are detectives.</p><p>They don’t just repeat patterns — they analyze clues.</p><p><strong>Metaphor:</strong></p><p>LLM = “Here are the words people usually say after this sentence.”</p><p>LRM = “Let me think this through logically and conclude what makes sense.”</p><p>It’s the difference between:</p><p>❌ <em>“Guess what comes next.”</em></p><p>✔️ <em>“Explain why this works.”</em></p><h3>3. Vector Database — A Library That Shelves Books by Meaning, Not Alphabet</h3><p>In a normal database, “apple” and “banana” are far apart because A ≠ B.</p><p>In a <strong>vector database</strong>, items are stored as mathematical vectors — basically coordinates that represent meaning.</p><p>So “apple,” “banana,” and “mango” would sit on the same shelf.</p><p><strong>Metaphor:</strong></p><p>A vector DB is like a library where books are arranged by <em>vibes</em> instead of titles.</p><p>Ask it for “fruits,” and it instantly pulls the nearest items in that semantic neighborhood.</p><h3>4. RAG (Retrieval-Augmented Generation) — The AI That Checks Its Work Before Speaking</h3><p>LLMs hallucinate. Everyone knows it.</p><p>RAG fixes that by letting the AI look things up in a vector DB <em>before</em> answering.</p><p><strong>Metaphor:</strong></p><p>Imagine ChatGPT with Google built-in.</p><p>Instead of saying, “I think I remember,” it says, “Hold on, let me check the reference.”</p><p>It retrieves facts → feeds them into the model → crafts a grounded answer.</p><p>It’s AI + research assistant.</p><h3>5. MCP (Model Context Protocol) — The Universal Power Adapter for AI Tools</h3><p>Every app, API, system, and tool has its own format.</p><p>MCP is a protocol that makes them all speak a <strong>common language</strong> so AI can use them easily.</p><p><strong>Metaphor:</strong></p><p>Think of MCP as a universal plug adapter:</p><ul><li>Want the AI to call your internal APIs? Plug it in.</li><li>Want it to talk to your database? Plug it in.</li><li>Want it to run scripts? Plug it in.</li></ul><p>AI becomes a “tool operator” instead of a text generator.</p><p>It turns models into <em>agents with superpowers</em>.</p><h3>6. MoE (Mixture of Experts) — Hiring Multiple Specialists Instead of One Superhuman</h3><p>Instead of one giant AI model doing everything, MoE breaks the work into specialists:</p><ul><li>one expert for math</li><li>one for language</li><li>one for coding</li><li>one for reasoning</li></ul><p>And a gatekeeper selects who handles which part of the question.</p><p><strong>Metaphor:</strong></p><p>MoE is like having a team of mini-experts inside a model —</p><p>you don’t need one person who knows everything; you need the right expert at the right time.</p><p>It saves compute and boosts intelligence.</p><h3>7. ASI (Artificial Super-Intelligence) — If the Entire Internet Became One Brain</h3><p>Right now we have:</p><ul><li>ANI — Narrow AI (specialized systems like Siri)</li><li>AGI — General AI (human-level intelligence; not here yet)</li><li><strong>ASI — Super-Intelligence (beyond any human; theoretical but talked about a lot)</strong></li></ul><p><strong>Metaphor:</strong></p><p>If AGI is “a smart person,”</p><p>ASI is “all of humanity’s intelligence combined into a single mind.”</p><p>It could learn, invent, innovate, and adapt far faster than we can.</p><p>It’s the AI equivalent of going from bicycles → rockets.</p><h3>Putting It All Together (A Story Version)</h3><p>If modern AI were a company:</p><ul><li><strong>Agentic AI</strong> is the proactive employee.</li><li><strong>LRM</strong> is the logical thinker who solves problems.</li><li><strong>Vector DB</strong> is the filing cabinet organized by meaning.</li><li><strong>RAG</strong> is the researcher who checks facts.</li><li><strong>MCP</strong> is the office communication system that connects all departments.</li><li><strong>MoE</strong> is the specialized team of experts.</li><li><strong>ASI</strong> is the mythical CEO from the future who can do everything better than humans — but hasn’t arrived yet.</li></ul><p>Learning AI starts making sense once you see how these pieces <em>fit together</em>, not as isolated buzzwords but as a system.</p><h3>Final Thought</h3><p>Understanding AI isn’t hard — it’s just badly explained most of the time.</p><p>Once I started translating jargon into metaphors, the fog lifted.</p><p>If you’re diving into AI, don’t memorize the terms.</p><p><strong>Relate them to things you already understand.</strong></p><p>That’s the fastest (and sanest) way to learn anything new.</p><p>It’s AI + research assistant.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ca447277b030" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data Transformations and Broadcasting in NumPy — How AI Handles Complex Data]]></title>
            <link>https://medium.com/@rageshramesh4/data-transformations-and-broadcasting-in-numpy-how-ai-handles-complex-data-05b2001486ce?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/05b2001486ce</guid>
            <category><![CDATA[numpy]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Sat, 22 Nov 2025 03:13:21 GMT</pubDate>
            <atom:updated>2025-11-22T03:13:21.704Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Data Transformations and Broadcasting in NumPy — How AI Handles Complex Data</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mou3ftNOeuexdNj8X8mlJQ.png" /></figure><p>If you’ve ever looked at NumPy code and thought,</p><p>“Wait… how did that tiny line transform my entire dataset?”</p><p>— trust me, you’re not alone.</p><p>I’ve been there, sitting with a cup of coffee, staring at a single NumPy operation, wondering how it just magically reshaped arrays like some kind of data yoga instructor. But once you understand <strong>data transformations</strong> and <strong>broadcasting</strong>, NumPy starts to look less like magic and more like… well, beautifully structured logic.</p><p>Today, let’s break it down in a calm, simple, friendly way — no stress, no complicated jargon, no frying your brain cells. Just pure clarity.</p><p>Ready? Let’s go.</p><h4>Why Data Transformations Matter in AI</h4><p>AI eats data. But here’s the catch: AI prefers its data in <em>specific shapes</em>.</p><p>- Images need to be 2D or 3D arrays.</p><p>- Batches of images need to be 4D arrays.</p><p>- Vectors need to be clean and consistent.</p><p>- Operations between arrays must match in shape.</p><p>So before an AI model ever “learns,” we humans have to gently massage the data into a form it understands. (Think of it like ironing clothes before a big event.)</p><p>That’s where <strong>transformations </strong>and <strong>broadcasting</strong> come in — they make messy data neat and compatible.</p><h3>Part 1: Data Transformations</h3><p>aka reshaping your messy room into something presentable</p><p>NumPy makes it easy to change the <a href="https://www.w3schools.com/python/numpy/numpy_array_shape.asp"><em>shape</em></a> of your data without changing the <a href="https://www.w3schools.com/python/numpy/numpy_creating_arrays.asp"><em>values</em></a>.</p><p>Let’s break down the key transformations:</p><h4>1. Reshaping — Same Data, New Shape</h4><p>Reshaping is like taking a long list and folding it nicely into rows and columns.</p><p><strong>Real-world analogy:</strong></p><p>You have 12 chocolates. You can arrange them in:</p><p>- 1 row of 12</p><p>- 2 rows of 6</p><p>- 3 rows of 4</p><p>Same chocolates. Different arrangement. 🍫</p><p><strong>Why AI cares:</strong></p><p>Most models need inputs in fixed shapes.</p><p>For example:</p><p>- An image might need to be shaped as <em>(height, width)</em></p><p>- A batch of images needs to be <em>(batch_size, height, width)</em></p><p>Reshaping prepares your data so models don’t complain.</p><h4>2. Flattening — Turning Anything Into a Line</h4><p>AI models (especially older ones) love working with <a href="https://en.wikipedia.org/wiki/Vector_(mathematics_and_physics)"><em>vectors</em></a>.</p><p>Flattening turns your matrix, cube, or giant 3D image into one long line.</p><p><strong>Why AI cares:</strong></p><p>The first layer of many neural networks expects a<em> </em><a href="https://www.geeksforgeeks.org/python/creating-a-one-dimensional-numpy-array/"><em>1D vector</em></a><em>.</em></p><p>Flattening gives that neat, aligned input.</p><h4>3. Expanding Dimensions — Adding Structure</h4><p>Sometimes AI frameworks want an extra <a href="https://en.wikipedia.org/wiki/Dimension"><em>dimension</em></a>:</p><p>- For channel information</p><p>- For batch size</p><p>- For compatibility</p><p>Expanding dimensions is like adding dividers into your suitcase.</p><p>Same stuff → more organized.</p><h3>Part 2: Broadcasting</h3><p>aka NumPy’s secret superpower</p><p>Broadcasting is the part everyone finds magical at first. And honestly… it is magical. But it has rules — simple ones, I promise.</p><h4>What Is Broadcasting?</h4><p>Broadcasting lets NumPy perform operations between arrays of <strong>different shapes</strong>, without demanding they be exactly the same.</p><p>Imagine you want to add:</p><p>- A full 3×3 matrix</p><p>- To a single number (like 5)</p><p>NumPy says: “No problem buddy, I’ll just stretch the 5 into a 3×3 behind the scenes.”</p><p>Or you want to subtract:</p><p>- A vector of shape (3,)</p><p>- From a matrix of shape (3, 3)</p><p>NumPy says: “I see what you’re going for. Let me line it up for you.”</p><blockquote>It’s like having an assistant who fills in the missing pieces without complaining.</blockquote><h4>Broadcasting Rule (The Only One You Need)</h4><p>NumPy compares shapes <strong>from right to left</strong>.</p><p>Two dimensions are compatible if:</p><p>- They are equal <em>or</em></p><p>- One of them is <em>1</em></p><p>If not compatible, NumPy refuses politely.</p><h4>Why AI Loves Broadcasting</h4><p>AI constantly does things like:</p><p>- Normalize data</p><p>- Scale pixel values</p><p>- Add biases to layers</p><p>- Combine feature maps</p><p>- Apply weights across batches</p><p>Broadcasting lets us apply these operations without writing loops.</p><p>(Loops are slow. AI hates slow.)</p><p><strong>Examples:</strong></p><p>- Dividing every pixel by 255</p><p>- Adding a bias vector to every row of a matrix</p><p>- Applying weights to data batches</p><p>All of this becomes clean, readable, and fast thanks to broadcasting.</p><h4>Real-World Analogy: Why Broadcasting Feels Magical</h4><p>Imagine you’re making sandwiches:</p><p>- You have 5 slices of bread</p><p>- One slice of cheese</p><p>Broadcasting is the friend who says:</p><p>“Relax, I’ll put that single slice of cheese on each sandwich for you.” You didn’t manually duplicate the cheese. They handled it.</p><p>NumPy does the same thing with values and arrays.</p><h3>Bringing It All Together (Calmly)</h3><p>Data transformations help you <strong>shape the data</strong>.</p><p>Broadcasting helps you <strong>operate on mismatched shapes without stress</strong>.</p><p>Together, they allow AI systems to:</p><p>- Handle batches elegantly</p><p>- Combine features effortlessly</p><p>- Apply operations across huge datasets</p><p>- Work faster and cleaner</p><p>And the best part?</p><p>Once you understand these two ideas, complex NumPy code suddenly becomes peaceful, predictable, and almost… fun.</p><p>Yes, I said it — NumPy can be fun.</p><p>If you’re enjoying this relaxed journey into AI (and not losing hair in the process), hit <strong>Follow.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=05b2001486ce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Chunking Changed the Way I Learn — Especially AI]]></title>
            <link>https://medium.com/@rageshramesh4/how-chunking-changed-the-way-i-learn-especially-ai-5301c4559e4d?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/5301c4559e4d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[growing]]></category>
            <category><![CDATA[chunking]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Tue, 11 Nov 2025 13:12:26 GMT</pubDate>
            <atom:updated>2025-11-11T13:12:26.998Z</atom:updated>
            <content:encoded><![CDATA[<h3>How Chunking Changed the Way I Learn — Especially AI</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1020/1*kps3FBDIyE2fB-hcXcec1Q.png" /></figure><p>Fifteen years into coding, I thought I had mastered the art of “just figuring it out.”</p><p>New framework? Sure. New library? Give me a weekend.</p><p>But when I started diving deep into <em>AI concepts</em>— linear algebra, neural networks, optimization — I hit a wall.</p><p>The problem wasn’t intelligence. It was <em>cognitive overload</em>.</p><p>Then I came across Barbara Oakley’s <em>A Mind for Numbers</em>, and one idea from that book rewired how I approach learning anything technical.</p><p>It’s called <strong>Chunking</strong>.</p><h4>The Moment It Clicked</h4><p>I was reading about gradient descent one night.</p><p>Every sentence made sense <em>individually</em>, but when I tried to connect them — loss functions, derivatives, partial gradients — it felt like trying to juggle spaghetti.</p><p>Then I stumbled upon Oakley’s explanation of <em>chunking</em>:</p><blockquote>A chunk is a compact package of information that your mind can easily access and use.</blockquote><p>It’s not about memorizing formulas.</p><p>It’s about <em>understanding and grouping small ideas</em> so they can be recalled and combined later.</p><p>That was the missing piece.</p><h4>What Chunking Really Means (In Developer Terms)</h4><p>Think of chunking like modular programming.</p><p>You don’t write one 5000-line function. You break it down into small, reusable units:</p><pre>function normalizeData(data) { /* … */ }<br>function computeGradient(weights, data) { /* … */ }<br>function updateWeights(weights, gradient) { /* … */ }</pre><p>Each function on its own doesn’t solve the problem.</p><p>But together, they make a working system.</p><p>Learning works the same way.</p><p>You create mental “functions” — chunks — that can be reused and combined.</p><h4>How I Started Chunking AI Concepts</h4><p>When I first tried to learn <strong>neural networks</strong>, I stopped reading whole chapters at once.</p><p>Instead, I built chunks:</p><p>1. <strong>Chunk 1 →</strong> <em>What’s a neuron?</em></p><p>Just a weighted sum + activation. That’s it.</p><pre>y = activation(w1*x1 + w2*x2 + b);</pre><p>2. <strong>Chunk 2 → </strong><em>What’s a layer?</em></p><p>A collection of neurons that process inputs in parallel.</p><p>3. <strong>Chunk 3 →</strong> <em>What’s backpropagation?</em></p><p>A method to update weights by flowing error backward.</p><blockquote>Once I could visualize each chunk independently, connecting them was effortless.</blockquote><blockquote>Suddenly, the neural network wasn’t a “mystical black box.” It was just three ideas talking to each other.</blockquote><h4>How to Practice Chunking</h4><p>Here’s the approach that worked for me (and I still use it whenever I hit something complex):</p><p><strong>1. Focus on one small concept at a time.</strong></p><p>No multitasking, no “just skimming.” Learn one idea deeply.</p><p><strong>2. Summarize in your own words.</strong></p><p>If you can explain it without notes, you’ve chunked it.</p><p><strong>3. Use spaced recall.</strong></p><p>Revisit your chunks after a day, a week, and a month. Each revisit strengthens the neural pathway.</p><p><strong>4. Connect old chunks to new ones.</strong></p><p>When learning “transformers” in AI, relate it to how RNNs process sequences. The bridge cements both.</p><h4>The Hidden Benefit: Confidence</h4><p>Before chunking, learning something new felt like drowning in an ocean of information. After chunking, it’s like walking across stepping stones. Each chunk is solid ground. You can always look back and say, “I know this much. I can build on it.”</p><h4>Final Thought</h4><p>In software development and AI, information changes fast.</p><p>Frameworks evolve, syntax changes, models get renamed.</p><p>But <strong>chunking</strong> — that skill of breaking, understanding, and rebuilding knowledge — stays relevant forever.</p><p>After 15 years of coding, I can confidently say: Learning isn’t about knowing more — it’s about remembering smarter. And chunking? That’s the best mental “algorithm” I’ve ever implemented</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5301c4559e4d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Silly Things Even Senior Developers Miss — My “unset” Moment in CSS]]></title>
            <link>https://medium.com/@rageshramesh4/silly-things-even-senior-developers-miss-my-unset-moment-in-css-cd7defb5f085?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd7defb5f085</guid>
            <category><![CDATA[senior-developer]]></category>
            <category><![CDATA[frontend]]></category>
            <category><![CDATA[css]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Tue, 04 Nov 2025 18:13:36 GMT</pubDate>
            <atom:updated>2025-11-04T18:13:36.661Z</atom:updated>
            <content:encoded><![CDATA[<h3>Silly Things Even Senior Developers Miss — My “unset” Moment in CSS</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/765/1*aTKBtaYCCReQE_BbbrDePA.png" /></figure><p>After 15 years of writing code across stacks — Node.js, React, AWS, and whatnot — I’d like to believe I’ve seen it all.</p><p>But every once in a while, something as innocent as a CSS property reminds me that humility is a crucial part of being a developer.</p><p>This is the story of how a single `unset` broke my entire desktop layout.</p><h3>The Setup</h3><p>I was tweaking a responsive layout in a React app. Our design system uses an array-based responsive syntax for styles. It looks something like this:</p><pre>&lt;Box<br>  width={[&#39;100%&#39;, &#39;50%&#39;, &#39;33%&#39;, &#39;25%&#39;]}<br>  margin={[&#39;8px&#39;, &#39;12px&#39;, null, null]}<br>/&gt;</pre><p>Everything was great — until I decided to “fix” a layout shift by resetting a margin at a specific breakpoint.</p><p>So I did this:</p><pre>&lt;Box<br>  width={[&#39;100%&#39;, &#39;50%&#39;, &#39;33%&#39;, &#39;25%&#39;]}<br>  margin={[&#39;8px&#39;, &#39;12px&#39;, &#39;unset&#39;, &#39;unset&#39;]}<br>/&gt;</pre><p>My logic was simple: “unset” means reset, right?</p><p>Wrong!</p><h3>The Unexpected Chaos</h3><p>On mobile and tablet screens, everything looked perfect.</p><p>But on desktop, things went haywire. Elements were overlapping, spacing vanished, and the layout suddenly looked like a CSS demolition derby.</p><p>For hours, I blamed the grid system, flexbox alignment, even the theme breakpoints.</p><p>Until I realized the villain was that one innocent-looking `unset`.</p><h3>The “Unset” Confusion</h3><p>Here’s the thing about CSS `unset`: It doesn’t mean <em>remove this property.</em></p><p>It means: <em>If this property is inherited, revert to the inherited value. Otherwise, set it to the initial (browser default) value.</em></p><p>For `margin`, which is not inherited, `<em>unset</em>` behaves like `<em>initial</em>`.</p><p>And the initial value of margin is 0.</p><p>So by saying <em>margin: unset</em>, I was effectively forcing margin: 0 on desktop breakpoints.</p><p>Here’s what I should have done:</p><pre>&lt;Box<br>width={[&#39;100%&#39;, &#39;50%&#39;, &#39;33%&#39;, &#39;25%&#39;]}<br>margin={[&#39;8px&#39;, &#39;12px&#39;, null, null]}<br>/&gt;</pre><p>By using `null`, the style library simply skips that breakpoint, leaving the previous value intact.</p><p>No explicit reset needed.</p><h3>The Lesson</h3><p>Even after 15 years, I still occasionally fall into the “I know what this means” trap.</p><p>CSS, for all its quirks, has a lot of these subtle gotchas:</p><ul><li>`unset` ≠ `remove`</li><li>`inherit` ≠ `default`</li><li>`initial` ≠ `browser default`</li></ul><p>And when you mix that with framework-level abstractions like responsive arrays, you can easily create a mess that only shows up on certain breakpoints.</p><h3>The Takeaway</h3><p>Don’t assume a property name means what it <em>sounds</em> like.</p><p>Before you use a keyword like `unset`, take 30 seconds to check the <a href="https://developer.mozilla.org/en-US/docs/Web/CSS/unset">docs</a>.</p><p>That 30 seconds can save you 3 hours of debugging.</p><p>Even after a decade and a half in full-stack development, I’ve learned that small mistakes like these don’t make us bad developers — they make us <em>human</em>.</p><p>And if you haven’t broken a production layout because of a single CSS keyword… just wait, your turn will come. 😅</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd7defb5f085" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Linear Algebra for AI — The Math Behind NumPy]]></title>
            <link>https://medium.com/@rageshramesh4/linear-algebra-for-ai-the-math-behind-numpy-815ecf470b37?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/815ecf470b37</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[ai-knowledge]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[numpy]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Sun, 02 Nov 2025 02:45:09 GMT</pubDate>
            <atom:updated>2025-11-02T02:45:09.029Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Linear Algebra for AI — The Math Behind NumPy</strong></h3><p>“At the heart of every AI model lies a simple truth — it’s all just math with direction.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j_7sERxHPK-1DArOgmiP2g.png" /></figure><h3>Why Learn Linear Algebra for AI?</h3><p>If statistics helps AI understand data, linear algebra helps it transform data.</p><p>Every image, sound, or sentence that AI processes is converted into numbers — and those numbers are stored as vectors and matrices.</p><p>Understanding how these structures work (and what we can do with them) gives you a peek under the hood of AI — the part that actually does the thinking.</p><h4>1. Vectors — Numbers With Direction</h4><p>Think of a vector as an arrow. It doesn’t just have a value — it has a direction and a magnitude.</p><p>Example:</p><p>- [2, 3] represents “go 2 steps right and 3 steps up”.</p><p>- In AI, this might be “feature 1 = height, feature 2 = weight”.</p><p>Each data point — an image pixel, a sound frequency, a word embedding — can be thought of as a vector.</p><p><strong>Why it matters in AI:</strong></p><ul><li>AI models use vectors to represent data inputs (like a row in your dataset).</li><li>Comparing two vectors helps measure <em>similarity</em> (used in recommendation systems and embeddings).</li><li>Operations on vectors help neural networks find *directions* that reduce error.</li></ul><h4>2. Matrices — The Engine of Transformation</h4><p>A matrix is just a grid of numbers — imagine a spreadsheet of values.</p><p>Example:</p><p>- A grayscale image can be represented as a 28×28 matrix of pixel values.</p><p>- A neural network layer can be seen as multiplying input vectors by a matrix of *weights*.</p><p>When we multiply a matrix by a vector, we’re transforming that vector — rotating it, scaling it, or projecting it into another dimension.</p><p><strong>Why it matters in AI:</strong></p><ul><li>Every layer in a neural network performs a <em>matrix multiplication</em>.</li><li>This operation transforms your input data into something more meaningful for the next layer.</li><li>The entire process of “learning” in AI is really about adjusting these matrix weights.</li></ul><h4>3. Matrix Multiplication — Combining Features</h4><p>Matrix multiplication is how we combine different features and weights. If a vector represents your input (say, [height, weight, age]) and a matrix represents the <em>model’s understanding</em> of how those inputs relate to an output, then multiplying them gives the <em>prediction</em>.</p><p><strong>AI connection:</strong></p><ul><li>This is literally how a neuron in a neural network computes its output.</li><li>Each neuron takes a weighted sum of its inputs — matrix multiplication does this efficiently.</li></ul><p>In other words, when you see an AI “predicting,” you’re watching <em>linear algebra in action.</em></p><h4>4. Identity and Inverse — Keeping Things Balanced</h4><p>The identity matrix acts like the number 1 in multiplication — multiply any matrix by it, and it stays the same. The inverse matrix, if it exists, is like dividing — it undoes a transformation.</p><p><strong>AI connection:</strong></p><ul><li>Optimization algorithms use concepts similar to matrix inverses when adjusting weights.</li><li>The idea of “reversing” a transformation helps neural networks correct their mistakes during training.</li></ul><h4>5. Eigenvalues and Eigenvectors — The Hidden Directions</h4><p>These sound scary, but here’s the intuition: Imagine spinning an object. Most points move, but some special directions don’t — they just stretch or shrink. Those special directions are <em>eigenvectors</em>, and the stretch amounts are <em>eigenvalues</em>.</p><p><strong>Why this matters in AI:</strong></p><ul><li>Techniques like <a href="https://www.geeksforgeeks.org/data-analysis/principal-component-analysis-pca/"><em>Principal Component Analysis (PCA)</em></a> use eigenvalues and eigenvectors to find the directions in your data with the most variation.</li><li>This helps AI reduce the number of features while keeping the most important information — making models faster and more efficient.</li></ul><h3>Bringing It All Together</h3><p>Linear algebra gives AI its structure — it’s the grammar that lets machines talk in the language of data.</p><p>When you:</p><ul><li>combine features → you’re doing <em>matrix multiplication</em></li><li>adjust model weights → you’re doing <em>vector updates</em></li><li>reduce dimensions → you’re using <em>eigenvectors and eigenvalues</em></li></ul><p>Every deep learning framework — TensorFlow, PyTorch, JAX — builds upon these same mathematical building blocks.</p><h4>Real-World Example — From Pixels to Predictions</h4><p>Let’s take an image of a handwritten “7” from the MNIST dataset:</p><p>1. The image (28×28 pixels) → a <em>matrix of numbers.</em></p><p>2. Flattened into a <em>vector</em> (784 values).</p><p>3. Multiplied by a <em>weight matrix</em> (learned by the model).</p><p>4. Transformed into <em>probabilities</em> of digits 0–9.</p><p>All of this is linear algebra — and the model learns by adjusting those matrices thousands of times per second.</p><h4>Why This Matters in NumPy</h4><p>When you use NumPy arrays, you’re not just manipulating data — you’re performing <em>the exact same operations AI models use internally</em>. NumPy is the bridge between math and machine learning. It turns these abstract ideas — vectors, matrices, transformations — into practical tools you can code.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=815ecf470b37" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Making Sense of Data: Mean, Median, Mode, Variance & Standard Deviation — The Math That Powers AI]]></title>
            <link>https://medium.com/@rageshramesh4/making-sense-of-data-mean-median-mode-variance-standard-deviation-the-math-that-powers-ai-ddca154767c8?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/ddca154767c8</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning-ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[mathematics]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Tue, 28 Oct 2025 17:43:46 GMT</pubDate>
            <atom:updated>2025-10-28T17:43:46.739Z</atom:updated>
            <content:encoded><![CDATA[<h3>Making Sense of Data: Mean, Median, Mode, Variance &amp; Standard Deviation — The Math That Powers AI</h3><blockquote><em>“Before machines can learn, they first need to understand what ‘normal’ looks like.”</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D8VDnDTV7Uqwx3SC2_mC0A.png" /></figure><h3>🧠 Why You Should Care About These Numbers</h3><p>When we talk about <em>Artificial Intelligence learning from data</em>, we often imagine something magical happening inside.</p><p>But at the core, AI isn’t guessing — it’s <strong>measuring</strong>.</p><p>It measures:</p><ul><li>What’s <em>normal</em> (average values)</li><li>How much things <em>vary</em></li><li>What <em>pattern</em> appears most often</li></ul><p>These five concepts — <strong>Mean, Median, Mode, Variance, and Standard Deviation</strong> — are the foundation of that measurement.</p><p>They help both humans and machines understand what kind of data we’re dealing with before making any predictions.</p><h3>🧩 1. Mean — Understanding the Center</h3><p>The <strong>mean</strong>, or average, gives you a single value that represents the <em>center</em> of your data.</p><p><strong>How it works:</strong></p><p>Add up all the numbers and divide by how many there are.</p><p><strong>Example:</strong></p><p>If five students scored 70, 80, 90, 100, and 60,</p><p>the mean is (70 + 80 + 90 + 100 + 60) / 5 = 80.</p><p>It tells us that, on average, students scored around 80.</p><h4>🧠 How AI Uses the Mean</h4><ol><li><strong>Data Normalization: </strong>Most machine learning models work better when data is centered around zero. To do that, we often subtract the <em>mean</em> from every value. This ensures all features (like height, age, income) are on similar scales, so no one feature dominates learning.</li><li><strong>Bias Detection: </strong>The mean can reveal hidden biases — for example, if the mean income in your dataset is much higher for one group than another, your model might unintentionally favor that group.</li><li><strong>Error Measurement: </strong>The concept of “average difference” (mean absolute error or mean squared error) is directly built on the mean — it tells the AI <em>how wrong</em> its predictions are on average.</li></ol><h3>⚖️ 2. Median — The Middle That Resists Outliers</h3><p>The <strong>median</strong> is the middle number when all values are arranged in order. It’s a great way to represent the “typical” value when some data points are extreme.</p><p><strong>Example:</strong></p><p>Imagine five salaries: ₹30k, ₹32k, ₹35k, ₹38k, ₹2,00,000.</p><p>The mean would be ₹67k, but the <strong>median</strong> is ₹35k — which better represents what most people earn.</p><h4>🧠 How AI Uses the Median</h4><ol><li><strong>Dealing with Outliers: </strong>Real-world data is messy. A few extreme values (like one very high or low number) can distort results. AI systems often use the <strong>median</strong> instead of the mean to detect or replace outliers without letting them affect learning.</li><li><strong>Model Robustness: </strong>Some algorithms (like median-based regression) prefer the median because it gives a more stable estimate when data isn’t perfectly distributed.</li><li><strong>Feature Scaling: </strong>In some preprocessing techniques, median values are used for imputation — filling missing data with a representative middle value.</li></ol><h3>3. Mode — The Most Common Pattern</h3><p>The <strong>mode</strong> is the value that appears most often in a dataset.</p><p><strong>Example:</strong></p><p>In a survey of favorite colors:</p><p>Blue, Blue, Green, Blue, Red, Green —</p><p>the mode is <em>Blue</em>.</p><p>It shows what’s most common, not necessarily what’s average.</p><h4>🧠 How AI Uses the Mode</h4><ol><li><strong>Classification Models: </strong>Many AI models, like decision trees and random forests, predict the most likely category — effectively choosing the <em>mode</em> of predictions from multiple trees or classifiers.</li><li><strong>Handling Categorical Data: </strong>When working with non-numerical data (like gender, city, or color), the mode helps identify the most frequent category for imputation or summarization.</li><li><strong>Ensemble Learning: </strong>In ensemble models, the final prediction is often the <em>mode</em> of all model predictions — this is called <strong>majority voting</strong>.</li></ol><h3>🔍 4. Variance — Measuring the Spread of Data</h3><p>Variance tells us <strong>how much the values differ from the mean</strong>. If data points are close to the mean, variance is small. If they’re widely spread out, variance is large.</p><p><strong>Example:</strong></p><p>Imagine two classrooms:</p><ul><li>In Class A, everyone scores between 75 and 85.</li><li>In Class B, scores range from 30 to 100.</li></ul><p>Both might have the same average score (say 80), but Class B has much <strong>higher variance</strong> — meaning students’ performances are all over the place.</p><h4>🧠 How AI Uses Variance</h4><ol><li><strong>Feature Importance: </strong>Features with higher variance often contain more useful information. If a feature barely changes, it might not help the model learn anything new.</li><li><strong>Dimensionality Reduction: </strong>Techniques like <strong>Principal Component Analysis (PCA)</strong> look for directions in the data with the highest variance — those are considered the most informative features for the model.</li><li><strong>Model Performance (Bias vs Variance Trade-off):</strong></li></ol><ul><li>In machine learning, variance isn’t just about data — it’s about models too.</li><li>A <strong>high-variance</strong> model fits training data too well but fails on new data (overfitting).</li><li>A <strong>low-variance</strong> model might underfit and miss important patterns.</li><li>Finding the right balance is a key part of training good AI models.</li></ul><h3>📈 5. Standard Deviation — Making the Spread Understandable</h3><p>Standard deviation is simply the <strong>square root of variance</strong>. While variance measures the spread in squared units, standard deviation brings it back to the same scale as the data — making it easier to interpret.</p><p><strong>Example:</strong></p><p>If the average temperature is 30°C and the standard deviation is 2°C, most days will fall between 28°C and 32°C — that’s how predictable the weather is.</p><h4>🧠 How AI Uses Standard Deviation</h4><ol><li><strong>Anomaly Detection: </strong>Points that are several standard deviations away from the mean are considered <em>anomalies</em>. For example, an AI fraud detection system might flag a transaction that’s way beyond the usual spending range.</li><li><strong>Model Confidence: </strong>In probabilistic AI models, standard deviation helps measure <strong>uncertainty</strong> — how confident the model is about a prediction. A small standard deviation means high confidence.</li><li><strong>Normalization (Z-score Scaling): </strong>Before training, data is often standardized using the below formula.</li></ol><pre>z = {x - mean}/standard deviation</pre><blockquote>This makes all features comparable by converting them to how many “standard deviations away” they are from the mean.</blockquote><h3>🌍 Real-World Analogy — How It All Connects</h3><p>Imagine building an AI to predict <strong>house prices</strong> in a city.</p><pre>Concept             | What It Reveals            | Example<br>--------------------|----------------------------|--------------------------------------------<br>Mean                | Average price              | ₹75 lakh — the overall trend<br>Median              | Typical price unaffected   | ₹60 lakh — what most homes actually cost<br>Mode                | Most common price bracket  | ₹50 lakh homes are most frequent<br>Variance            | How spread prices are      | Tells how uniform or diverse the market is<br>Standard Deviation  | Predictability of prices   | Small SD → stable market, large SD → volatile</pre><p>Before training your model, understanding these five tells you:</p><ul><li>whether your data is clean</li><li>if it’s biased</li><li>how much variation the model needs to handle.</li></ul><p>This is why <strong>data analysis is the first step before AI training ever begins</strong>.</p><h3>🧭 Bringing It All Together</h3><ul><li><strong>Mean</strong> → helps find the center of data and normalize it</li><li><strong>Median</strong> → protects your model from extreme outliers</li><li><strong>Mode</strong> → captures the most common class or behavior</li><li><strong>Variance</strong> → tells you how unpredictable your data is</li><li><strong>Standard Deviation</strong> → gives you a practical sense of that unpredictability</li></ul><p>Together, these five are the <strong>language of data </strong>and learning them helps you speak the same language as your AI models.</p><h3>✨ Stay Updated</h3><p>If you found this helpful and want to keep learning <strong>AI and ML from the ground up</strong>, follow me here on <strong>Medium.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ddca154767c8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[NumPy for AI Beginners: How Machines Understand Numbers.]]></title>
            <link>https://medium.com/@rageshramesh4/numpy-for-ai-beginners-how-machines-understand-numbers-6bb52a2ca673?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/6bb52a2ca673</guid>
            <category><![CDATA[numpy]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Sat, 25 Oct 2025 04:50:07 GMT</pubDate>
            <atom:updated>2025-11-22T03:14:19.325Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><em>“AI starts with numbers — and NumPy is how we teach computers to handle them.”</em></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*47Bd_V3S89sFdj5smf7-QQ.png" /><figcaption>NumPy for AI</figcaption></figure><h3>🧠 Why NumPy Matters in AI</h3><p>Before we dive into neural networks or fancy algorithms, we need to answer a simple question:</p><p>👉 <strong>How does a computer understand and process data?</strong></p><p>Everything — an image, a word, a sound — is eventually represented as <strong>numbers</strong>.</p><p>For AI to work, it needs a powerful way to <strong>store, manipulate, and compute</strong> those numbers efficiently.</p><p>That’s where <strong>NumPy (Numerical Python)</strong> comes in.</p><p>Think of NumPy as <strong>the mathematical engine</strong> behind every AI and data science library — TensorFlow, PyTorch, Scikit-learn — they all rely on NumPy under the hood.</p><h3>💡 What Is NumPy, Really?</h3><p>NumPy is a Python library built for <strong>fast numerical computation</strong>.</p><p>It introduces a special data type called a <strong>NumPy array</strong>, which is like a <strong>supercharged list</strong>.</p><p><strong>Why not just use Python lists?</strong></p><p>Because lists are:</p><ul><li>Slower</li><li>Memory inefficient</li><li>Can’t handle large matrix operations easily</li></ul><p><strong>NumPy arrays, on the other hand:</strong></p><ul><li>Are <strong>fast</strong> (written in C under the hood)</li><li>Support <strong>vectorized operations</strong></li><li>Work seamlessly with <strong>matrices</strong>, <strong>vectors</strong>, and <strong>multi-dimensional data</strong></li></ul><h3>🧩 Installing and Importing NumPy</h3><pre>pip install numpy</pre><p>import it:</p><pre>import numpy as np</pre><p>You’ll see np used everywhere — it’s the standard alias.</p><h3>🔢 Arrays: The Heart of NumPy</h3><p>Let’s start with something simple:</p><pre>import numpy as np<br><br>a = np.array([1, 2, 3])<br>print(a)<br># Output: [1 2 3]</pre><p>Looks similar to a Python list, right? But here’s the magic:</p><pre>print(a + 5)<br># Output: [6 7 8]</pre><p>That single line adds <strong>5 to every element</strong> — automatically.In a normal Python list, you’d need a loop. NumPy does it in one shot using <strong>vectorization</strong> — this is key for AI, where we often deal with millions of numbers at once.</p><p>Think of <strong>vectorization</strong> as teaching your computer to <strong>do more in one go</strong> — instead of doing things one by one.</p><p><strong>🐢 The Old Way (Using Loops)</strong></p><pre>numbers = [1, 2, 3, 4, 5]<br>result = []<br>for n in numbers:<br>    result.append(n + 10)<br><br>print(result)<br># [11, 12, 13, 14, 15]</pre><blockquote>This works — but it’s slow. The computer is adding one number at a time.</blockquote><p><strong>🚀 The NumPy Way (Vectorization)</strong></p><pre>import numpy as np<br><br>numbers = np.array([1, 2, 3, 4, 5])<br>result = numbers + 10<br>print(result)<br># [11 12 13 14 15]</pre><blockquote>That single line replaces the entire loop! NumPy applies the operation to <strong>all elements at once</strong> — like a turbocharged calculator.</blockquote><h3>🧮 Basic Operations That Power AI</h3><p>Let’s explore a few operations you’ll use constantly:</p><h4>1. Matrix Addition</h4><pre>a = np.array([[1, 2], [3, 4]])<br>b = np.array([[5, 6], [7, 8]])<br>print(a + b)<br># [[ 6  8]<br>#  [10 12]]</pre><h4>2. Matrix Multiplication</h4><pre>np.dot(a, b)<br># [[19 22]<br>#  [43 50]]</pre><p>This is used in every <strong>neural network layer</strong> — it’s how inputs get transformed by weights.</p><h4>3. Element-wise Operations</h4><pre>print(a * 2)<br># [[2 4]<br>#  [6 8]]</pre><h4>4. Statistical Functions</h4><pre>data = np.array([2, 4, 6, 8, 10])<br>print(np.mean(data))   # Average<br>print(np.std(data))    # Standard deviation</pre><p>These help measure things like <strong>loss</strong>, <strong>accuracy</strong>, or <strong>data normalization</strong>.</p><h3>⚙️ Real AI Example: Normalizing Data</h3><p>Before training an AI model, we often need to <strong>normalize</strong> data — scale all values between 0 and 1.</p><pre>pixels = np.array([0, 128, 255])<br>normalized = pixels / 255<br>print(normalized)<br># Output: [0.         0.50196078 1.        ]</pre><p>This makes training more stable and ensures that no feature dominates due to scale differences.</p><h3>🧭 Real-World Analogy</h3><p>Imagine you’re managing a restaurant with thousands of daily orders.</p><ul><li>With <strong>Python lists</strong>, you’d manually check each order one by one.</li><li>With <strong>NumPy</strong>, you process all orders in one go — like a conveyor belt that calculates totals, applies discounts, and predicts demand automatically.</li></ul><p>That’s the kind of <strong>parallel thinking</strong> NumPy brings to AI.</p><h3>🧠 Why AI Engineers Love NumPy</h3><p>NumPy is like the <strong>grammar</strong> of AI — once you know it, everything else makes sense.</p><p>Libraries like TensorFlow and PyTorch are built on the same principles:</p><ul><li>Arrays (tensors)</li><li>Vectorized math</li><li>Matrix operations</li></ul><p>If you master NumPy, you’ll understand how <strong>AI computes</strong> — not just what it computes.</p><h3>✍️ Concepts Recall</h3><pre>| Concept                  | Example               | Why It Matters                 |<br>|-----------------------   |-----------------------|--------------------------------|<br>| **NumPy Arrays**         | `np.array([1,2,3])`   | Foundation of AI math          |<br>| **Vectorization**        | `a + 5`               | Speeds up computations         |<br>| **Matrix Multiplication**| `np.dot(a,b)`         | Core of neural networks        |<br>| **Normalization**        | `pixels / 255`        | Makes training stable          |<br>| **Statistical Functions**| `np.mean(), np.std()` | Used in loss and analysis  |</pre><h3>🚀 What’s Next</h3><p>In the next article, we’ll explore the follow ups on Numpy to go deeper and understand main concepts</p><ul><li><a href="https://medium.com/@rageshramesh4/linear-algebra-for-ai-the-math-behind-numpy-815ecf470b37">Linear Algebra for AI — The Math Behind NumPy”</a></li><li><a href="https://medium.com/@rageshramesh4/data-transformations-and-broadcasting-in-numpy-how-ai-handles-complex-data-05b2001486ce?postPublishedType=initial">Data Transformations and Broadcasting in NumPy — How AI Handles Complex Data”</a></li><li>Vectorization — How to Make Python Blazingly Fast Without Writing a Single Loop</li><li>The Math You Actually Need for AI — Explained with NumPy</li></ul><blockquote><em>🧠 </em>Follow me on Medium for more hands-on AI articles published every Saturday!</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6bb52a2ca673" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learn AI from Scratch: The Roadmap I Wish I Had]]></title>
            <link>https://medium.com/@rageshramesh4/learn-ai-from-scratch-the-roadmap-i-wish-i-had-2b9fc44ca8b9?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b9fc44ca8b9</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[roadmaps]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Tue, 21 Oct 2025 09:24:22 GMT</pubDate>
            <atom:updated>2025-10-25T04:52:35.737Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1017/1*Tg4K2PqBnK4-Zk2QkdH93g.png" /></figure><h3>Why This Series Exists</h3><p>Artificial Intelligence (AI) can feel like an exclusive club — filled with math, jargon, and intimidating words like “backpropagation,” “gradient descent,” or “transformers.”</p><p>But here’s the truth: <strong>you don’t need a PhD, a fancy laptop, or years of research to start learning AI.</strong></p><p>All you need is:</p><ul><li>Basic Python knowledge 🐍</li><li>High school-level math 🧮</li><li>And a curious mind 🤔</li></ul><p>That’s it.</p><p>This Medium series is for anyone who’s ever thought:</p><blockquote><em>“I want to learn AI, but I don’t know where to start.”</em></blockquote><h3>What You’ll Learn</h3><p>Starting this week, I’ll be publishing <strong>two articles every week (Thursday &amp; Saturday)</strong> that take you <strong>from zero to AI practitioner</strong> — step by step, no shortcuts, no fluff.</p><p>Here’s what’s coming up:</p><h4>🔹 The Building Blocks</h4><ul><li><strong>NumPy</strong> — how AI understands numbers</li><li><strong>Pandas</strong> — how we feed AI real-world data</li><li><strong>Matplotlib</strong> — how to see what your AI sees</li></ul><h4>🔹 The Power Tools</h4><ul><li><strong>TensorFlow</strong> — the engine that runs deep learning</li><li><strong>PyTorch</strong> — a more flexible, pythonic approach</li><li><strong>JAX</strong> — high-performance computing made simple</li></ul><h4>🔹 The Practice Ground</h4><ul><li><strong>Kaggle</strong> — your playground for real datasets</li><li><strong>Mini Projects</strong> — simple AI models that <em>actually work</em></li></ul><h4>🔹 The Math You’ll Need (and nothing more)</h4><p>We’ll cover just enough math to understand what’s going on behind the scenes:</p><ul><li>Linear Algebra (vectors, matrices)</li><li>Probability &amp; Statistics</li><li>Calculus (for optimization intuition)</li></ul><h4>🔹 The Big Picture</h4><p>Once you’ve mastered the tools and math, we’ll move to:</p><ul><li><strong>Machine Learning</strong> — teaching computers to learn patterns</li><li><strong>Deep Learning</strong> — building neural networks like the ones behind ChatGPT and Stable Diffusion</li></ul><h3>The Goal</h3><p>By the end of this series, you’ll:</p><ul><li>Understand how modern AI models are built and trained</li><li>Be able to work with real datasets</li><li>Recreate simple ML and DL projects from scratch</li><li>Have a solid foundation to build your own AI projects</li></ul><p>This isn’t about shortcuts.</p><p>This is about <strong>understanding AI the right way — from the ground up.</strong></p><h3>What You Can Do Now</h3><ol><li><strong>Follow me here on Medium</strong> so you don’t miss new posts.</li><li><strong>Comment what you’d like to learn first</strong> — I might reorder topics based on your feedback.</li><li><strong>Share this with your friends or colleagues</strong> who’ve been curious about AI but didn’t know where to begin.</li></ol><h3>Final Thoughts</h3><p>AI isn’t magic — it’s math, code, and creativity working together. If you can write a few lines of Python and remember high school algebra, you can absolutely learn AI. Let’s make this journey together — from “I don’t understand AI” to “I just built my first neural network.”</p><p><strong>Start out with : “</strong><a href="https://medium.com/@rageshramesh4/numpy-for-ai-beginners-how-machines-understand-numbers-6bb52a2ca673"><strong>NumPy for AI Beginners.</strong></a><strong>”</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b9fc44ca8b9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Multiple language support in Lambdas using layers.]]></title>
            <link>https://awstip.com/multiple-language-support-in-lambdas-using-layers-6c34f31d802f?source=rss-57ba6bf448e------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c34f31d802f</guid>
            <category><![CDATA[aws-lambda-layer]]></category>
            <category><![CDATA[lambda]]></category>
            <category><![CDATA[lambda-function]]></category>
            <category><![CDATA[nodejs]]></category>
            <category><![CDATA[lambda-layer]]></category>
            <dc:creator><![CDATA[Ragesh R]]></dc:creator>
            <pubDate>Sat, 09 Apr 2022 05:49:20 GMT</pubDate>
            <atom:updated>2022-04-11T07:49:13.527Z</atom:updated>
            <content:encoded><![CDATA[<p>Lambdas are compute services that lets you run code without provisioning or managing servers. They are individual containers that run specific code without any relation to other lambdas or other instances of the same lambda.</p><p>But what if a scenario arises where we need to share some code between lambdas, and we do not want to rewrite code for all lambdas when there is a change in this specific code.</p><p>Thats were layers come in. Lambda <a href="https://docs.aws.amazon.com/lambda/latest/dg/gettingstarted-concepts.html#gettingstarted-concepts-layer">layers</a> provide a convenient way to package libraries and other dependencies that you can use with your Lambda functions. This helps in reducing lambda size as well as sharing.</p><h3>Requirement</h3><p>Implement internalisation for lambda functions. There are about 10 lambda functions each returning anywhere from 5–15 different texts in responses. We will need to internalise these texts so that depending on user locale we show the relevant messages.</p><h3>Traditional way of doing it</h3><p>Using i18n npm package we can internalize the strings like below and add all these codes to all the lambdas.</p><ol><li>Install <a href="https://www.npmjs.com/package/i18n">i18n</a></li></ol><pre>npm install i18n --save</pre><p>2. Setup configuration for i18n</p><pre>const i18n = require(&#39;i18n&#39;);<br>const path = require(&#39;path&#39;);<br>i18n.configure({<br>  locales: [&#39;en&#39;, &#39;el&#39;],<br>  defaultLocale: &#39;en&#39;,<br>  queryParameter: &#39;lang&#39;,<br>  directory: path.join(__dirname, &#39;locales&#39;),<br>  api: {<br>    &#39;__&#39;: &#39;translate&#39;,  <br>    &#39;__n&#39;: &#39;translateN&#39; <br>  },<br>  objectNotation: true<br>});<br>module.exports =  i18n;</pre><p>3. Setup local service class that helps in text translation</p><pre>/**<br> * LocaleService<br> */<br>class LocaleService {<br>    /**<br>     *<br>     * <a href="http://twitter.com/param">@param</a> i18nProvider The i18n provider<br>     */<br>    constructor(i18nProvider, locale) {<br>      this.i18nProvider = i18nProvider;</pre><pre>if (locale &amp;&amp; this.getLocales().indexOf(locale) !== -1) {<br>        this.i18nProvider.setLocale(locale)<br>      }<br>    }<br>    /**<br>     *<br>     * <a href="http://twitter.com/returns">@returns</a> {string} The current locale code<br>     */<br>    getCurrentLocale() {<br>      return this.i18nProvider.getLocale();<br>    }<br>    /**<br>     *<br>     * <a href="http://twitter.com/returns">@returns</a> string[] The list of available locale codes<br>     */<br>    getLocales() {<br>      return this.i18nProvider.getLocales();<br>    }<br>    /**<br>     *<br>     * <a href="http://twitter.com/param">@param</a> string String to translate<br>     * <a href="http://twitter.com/param">@param</a> args Extra parameters<br>     * <a href="http://twitter.com/returns">@returns</a> {string} Translated string<br>     */<br>    translate(string, args = undefined) {<br>      return this.i18nProvider.__(string, args)<br>    }</pre><pre>}</pre><pre>module.exports = LocaleService</pre><p>4. Setup locales folder with language json files like below</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/684/1*2J0PKZkyKGS2IKZZDitl-g.png" /><figcaption>Locales folder</figcaption></figure><p>5. To use the locale service and translate string we will do the below</p><pre>const LocaleService = require(&quot;./localeService&quot;);<br>const i18n = require(&quot;./i18n.config.js&quot;);</pre><pre>const locale = event.queryStringParameters.locale</pre><pre>const localeService = new LocaleService(i18n, locale);<br> <br>console.log(localeService.translate(&quot;generic.requestInvalid&quot;))</pre><blockquote>Here the localeService.translate(&quot;generic.requestInvalid&quot;) will translate the string present inside the relevant json file as per the passed in locale.</blockquote><p>The above flow works fine but what happens if we need to change some texts in the json file. Let’s say the same code is replicated in 10 different lambdas. So each change should be done across 10 different lambdas and updated.</p><h3>Modularisation</h3><p>For modularisation we will be using layers in lambda.</p><ol><li>Create a new node project with the parent folder as <strong>nodejs</strong> (This file name cannot be changed as this is the naming convention for nodejs module in layers). Inside this nodejs file we will move all our internalisation files as below. The nodejs.zip file in the below image is the zip file we get when we zip our nodejs parent folder.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/688/1*aerAy0bw_GTwuiEKlCF8qQ.png" /><figcaption>Project structure</figcaption></figure><p>2. In the aws lambda console we have an option as layers in left hand side. We click on that and create a new layer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aukSPeA4y6rlS8tjYdMp3Q.png" /><figcaption>Layer in aws lambda dashboard</figcaption></figure><p>3. While creating layer. We need to give the layer name, compatible architectures and compatible run time. We will also need to upload a zip file. In our case the nodejs.zip</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Yvpu7I4mG9CPdPQ-6Ho6Hg.png" /><figcaption>Layer creation</figcaption></figure><p>4. Once the layer is created we can attach it to lambdas directly from the lambda function overview.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5QhS0EYOqC8DSRoJ3dIl1A.png" /><figcaption>Layer attached to lambda</figcaption></figure><p>5. In the lambda, we can require any module inside the layer using the /opt/ path like below</p><pre>const LocaleService = require(&quot;/opt/localeService&quot;);<br>const i18n = require(&quot;/opt/i18n.config.js&quot;);</pre><p>Layers can also be used to reduce lambda size. We can move heavy, rarely changing packages to layers so that every time we change anything in lambda we do not need to upload the whole code again.</p><p>If you liked this story please clap and follow for interesting articles every week.</p><p>Happy Coding!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c34f31d802f" width="1" height="1" alt=""><hr><p><a href="https://awstip.com/multiple-language-support-in-lambdas-using-layers-6c34f31d802f">Multiple language support in Lambdas using layers.</a> was originally published in <a href="https://awstip.com/">AWS Tip</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>