<?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 Akshay Adlakha on Medium]]></title>
        <description><![CDATA[Stories by Akshay Adlakha on Medium]]></description>
        <link>https://medium.com/@akshayadlakha1995?source=rss-f9dfd297d92------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*g13AOA-7BWlPXqP39oEzcg.jpeg</url>
            <title>Stories by Akshay Adlakha on Medium</title>
            <link>https://medium.com/@akshayadlakha1995?source=rss-f9dfd297d92------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 01:30:17 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@akshayadlakha1995/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[The ML Model Is Not Dead — It Is Becoming the Intelligence Layer of AI Systems]]></title>
            <link>https://medium.com/codetodeploy/the-ml-model-is-not-dead-it-is-becoming-the-intelligence-layer-of-ai-systems-cf760c681f30?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf760c681f30</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Fri, 24 Jul 2026 00:04:38 GMT</pubDate>
            <atom:updated>2026-07-25T15:08:59.827Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>Why the future of enterprise AI isn’t about replacing machine learning — it’s about turning models into reusable intelligence services that power reasoning, prediction, and decision-making.</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UMeA-SGIUdELj3drm1Lmkg.png" /></figure><blockquote><strong><em>💥 Master Any Skills in 3 Months</em></strong><br><em> </em>📚 Up to <strong>50% OFF</strong> Premium Courses<br> ⏰ Limited-Time Offer<br><a href="https://trk.udemy.com/zz4NBO"><em>👉 </em><strong><em>Enroll Now &amp; Start Learning</em></strong></a></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pkQFf0NUnzoFbVEVQCF5WQ.png" /></figure><p>For years, building AI systems followed a familiar pattern.</p><p>Collect data.</p><p>Train a model.</p><p>Measure accuracy.</p><p>Deploy the model.</p><p>The model was the product.</p><p>A recommendation model powered recommendations.</p><p>A fraud model detected fraud.</p><p>A forecasting model predicted demand.</p><p>The machine learning model sat at the center of the application.</p><p>But something changed.</p><p>The rise of foundation models and generative AI created a narrative:</p><p><strong>“Traditional machine learning is being replaced.”</strong></p><p>Teams started asking:</p><ul><li>Do we still need XGBoost?</li><li>Are neural networks becoming irrelevant?</li><li>Will LLMs replace predictive models?</li><li>Is classical ML becoming legacy technology?</li></ul><p>The answer is more nuanced.</p><p>The ML model is not disappearing.</p><p>It is changing roles.</p><p>The future AI system will not be a single giant model answering every question.</p><p>It will be an ecosystem where multiple intelligence components work together.</p><p>And traditional ML models will become one of the most important layers inside that ecosystem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Rh7BsAwyMLHWN3fi9ZmAJQ.png" /><figcaption><strong>The Evolution of AI Systems</strong></figcaption></figure><p><strong>The Model Was Never the Product</strong></p><p>A common misunderstanding is that the AI model itself creates business value.</p><p>It does not.</p><p>The model is only one component.</p><p>A credit risk model with 95% accuracy does not create value unless it can:</p><ul><li>evaluate applications,</li><li>explain decisions,</li><li>integrate with workflows,</li><li>reduce risk,</li><li>improve customer experience.</li></ul><p>The same applies today.</p><p>An LLM by itself is not an enterprise AI system.</p><p>A GPT-style model can generate text, but it does not know:</p><ul><li>Which customer should receive an offer,</li><li>Which manufacturing cell will fail tomorrow,</li><li>Which transaction is fraudulent,</li><li>Which vehicle component is degrading,</li><li>Which supplier creates quality risk?</li></ul><p>These require specialized intelligence.</p><p>And that is where ML models become critical.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L6-Ks1lWJBb3en3i5wrs3g.png" /><figcaption><strong>AI System Stack</strong></figcaption></figure><p><strong>LLMs Are General Intelligence — ML Models Are Specialized Intelligence</strong></p><p>One of the biggest misconceptions in AI today is:</p><p>“One powerful model will solve everything.”</p><p>In reality, intelligence has different forms.</p><p>Humans have:</p><ul><li>general reasoning ability,</li><li>domain expertise,</li><li>memory,</li><li>intuition,</li><li>specialized skills.</li></ul><p>AI systems will evolve similarly.</p><p>Large language models provide broad reasoning capability.</p><p>But specialized ML models provide precise decision-making.</p><p>Consider manufacturing.</p><p>An LLM can answer:</p><blockquote>“Explain why this battery cell might fail.”</blockquote><p>But a predictive ML model can answer:</p><blockquote>“This cell has an 87% probability of failure during formation because ACIR variance increased, coating thickness changed, and temperature behavior deviated.”</blockquote><p>The LLM explains.</p><p>The ML model predicts.</p><p>Together, they create intelligence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hepEBQzIdiJzT35jMYnsIw.png" /><figcaption><strong>LLM + ML Model Collaboration</strong></figcaption></figure><p><strong>The Future Is Not LLM vs ML — It Is LLM + ML</strong></p><p>The industry discussion has become a competition:</p><p>Traditional ML vs Generative AI.</p><p>But enterprises do not operate that way.</p><p>They need systems.</p><p>A modern AI architecture will look more like:</p><p><strong>Example: Automotive Battery Manufacturing AI Assistant</strong></p><p>A manufacturing engineer asks:</p><blockquote>“Why did today’s battery production batch have higher rejection rates?”</blockquote><p>The AI system performs multiple tasks:</p><p><strong>Step 1: LLM understands the question</strong></p><p>“What is the engineer asking?”</p><p><strong>Step 2: The retrieval system gathers context</strong></p><ul><li>production logs</li><li>quality reports</li><li>sensor data</li><li>maintenance history</li></ul><p><strong>Step 3: ML models analyze signals</strong></p><p>Failure prediction model:</p><p>Probability of defect = 82%</p><p>Anomaly detection model:</p><p>Coating thickness deviation detected</p><p>Computer vision model:</p><p>Visual defect identified</p><p><strong>Step 4: LLM generates an explanation</strong></p><p>“The rejection increase is primarily associated with coating variation in Line 3 during the afternoon shift.”</p><p>The intelligence did not come solely from the LLM.</p><p>It came from multiple models working together.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PMxXNCRrx6Jdkf6HTV5BMg.png" /><figcaption><strong>Enterprise AI Agent Architecture</strong></figcaption></figure><p><strong>ML Models Are Becoming the “Sensors” of AI Systems</strong></p><p>A useful analogy:</p><p>LLMs are like the brain.</p><p>But brains need senses.</p><p>ML models provide those senses.</p><p>A human does not understand the world only through reasoning.</p><p>They need:</p><ul><li>vision,</li><li>hearing,</li><li>memory,</li><li>specialized skills.</li></ul><p>AI systems work similarly.</p><p>Examples:</p><p><strong>Computer Vision Models</strong></p><p>They detect:</p><ul><li>manufacturing defects,</li><li>safety violations,</li><li>vehicle damage.</li></ul><p><strong>Forecasting Models</strong></p><p>They predict:</p><ul><li>demand,</li><li>inventory,</li><li>energy consumption.</li></ul><p><strong>Classification Models</strong></p><p>They identify:</p><ul><li>fraud,</li><li>customer churn,</li><li>risk.</li></ul><p><strong>Recommendation Models</strong></p><p>They optimize:</p><ul><li>products,</li><li>content,</li><li>next actions.</li></ul><p>These models become intelligence modules that AI systems consume.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xKDMYm9yh8hTElq3OLuHtQ.png" /></figure><p><strong>Why Small Specialized Models Will Become More Valuable</strong></p><p>There is another important shift happening.</p><p>The future will not be about bigger models alone.</p><p>It will also be about better orchestration.</p><p>A trillion-parameter model is expensive.</p><p>Running it for every decision is inefficient.</p><p>Companies need:</p><ul><li>lower latency,</li><li>lower cost,</li><li>predictable behavior,</li><li>explainability.</li></ul><p>A small ML model can often outperform an LLM for specific tasks.</p><p>Example:</p><p>Question:</p><blockquote>“Will this battery cell fail?”</blockquote><p>A specialized XGBoost model trained on millions of historical battery measurements may outperform an LLM.</p><p>Because the problem is not language.</p><p>It is a prediction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BoH-UktIxexgczjXIi2HxA.png" /><figcaption><strong>Intelligent Model Routing</strong></figcaption></figure><p><strong>The New Role of Data Scientists</strong></p><p>The role of data scientists is changing.</p><p>Previously:</p><blockquote>Build the best model.</blockquote><p>Future:</p><blockquote>Build the intelligence components that enable AI systems.</blockquote><p>The questions change.</p><p>Old questions:</p><ul><li>Which algorithm gives the highest accuracy?</li><li>Which features improve AUC?</li><li>Which model wins the benchmark?</li></ul><p>New questions:</p><ul><li>Where does prediction belong in the AI workflow?</li><li>When should the system call an ML model?</li><li>How should models communicate with agents?</li><li>How do we evaluate the complete AI system?</li></ul><p>The model is no longer the final product.</p><p>It is a capability inside a larger intelligence architecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PdaO6qI0qE1tfBUOMR1Xqw.png" /></figure><p><strong>The Future Enterprise AI Stack</strong></p><p>The winning AI architectures will combine:</p><p><strong>Foundation Models</strong></p><p>For:</p><ul><li>reasoning</li><li>language understanding</li><li>planning</li></ul><p><strong>Machine Learning Models</strong></p><p>For:</p><ul><li>prediction</li><li>optimization</li><li>scoring</li></ul><p><strong>Retrieval Systems</strong></p><p>For:</p><ul><li>enterprise knowledge</li><li>context</li></ul><p><strong>Rules &amp; Policies</strong></p><p>For:</p><ul><li>governance</li><li>compliance</li></ul><p><strong>Evaluation Systems</strong></p><p>For:</p><ul><li>reliability</li><li>monitoring</li></ul><p>No single component replaces another.</p><p>They amplify each other.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oWvlBSZZiHWG2OS_z2s4IQ.png" /></figure><p><strong>The Model Is Moving Up the Stack</strong></p><p>The biggest mistake organizations can make is thinking:</p><p>“Generative AI replaced machine learning.”</p><p>The reality is:</p><p>Generative AI changed where machine learning fits.</p><p>The ML model is no longer the entire application.</p><p>It is becoming the intelligence layer that gives AI systems:</p><ul><li>prediction,</li><li>precision,</li><li>specialization,</li><li>decision-making capability.</li></ul><p>The future does not belong to LLM-only systems.</p><p>It belongs to systems where:</p><p><strong>LLMs think.<br>ML models predict.<br>Retrieval provides context.<br>Agents execute actions.<br>Evaluation ensures trust.</strong></p><p>The next generation of AI will not be built around one model.</p><p>It will be built around intelligent systems.</p><p>And ML models will remain at the center of that evolution.</p><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d9QTaaaxboQP_gKSLedW_w.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://www.linkedin.com/in/bhuwan-kc-583505319/?isSelfProfile=true"><strong>Linkedin</strong></a>| <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p><strong>Disclosure:</strong> This post includes affiliate and partnership links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf760c681f30" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/the-ml-model-is-not-dead-it-is-becoming-the-intelligence-layer-of-ai-systems-cf760c681f30">The ML Model Is Not Dead — It Is Becoming the Intelligence Layer of AI Systems</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Every AI Product Is Secretly a Search Engine]]></title>
            <link>https://medium.com/codetodeploy/why-every-ai-product-is-secretly-a-search-engine-227495efa4a4?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/227495efa4a4</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 22:07:33 GMT</pubDate>
            <atom:updated>2026-08-03T18:26:35.531Z</atom:updated>
            <content:encoded><![CDATA[<p><strong><em>The smartest AI systems don’t know everything — they know how to find the right information at the right time.</em></strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tvApJxkVhaK1dsGIsZ61Bg.png" /></figure><p>When ChatGPT answers your question in seconds, it feels almost magical.</p><p>You ask.</p><p>It responds.</p><p>The conversation flows naturally, as if the model already knows everything it needs.</p><blockquote><strong><em>💥 Master Any Skills in 3 Months</em></strong><br><em> </em>📚 Up to <strong>50% OFF</strong> Premium Courses<br> ⏰ Limited-Time Offer<br><a href="https://trk.udemy.com/zz4NBO"><em>👉 </em><strong><em>Enroll Now &amp; Start Learning</em></strong></a></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pkQFf0NUnzoFbVEVQCF5WQ.png" /></figure><p>That’s the illusion.</p><p>Behind nearly every production AI application is a hidden system working furiously before the model generates its first word.</p><p>Documents are searched.</p><p>Past conversations are retrieved.</p><p>Business rules are loaded.</p><p>Databases are queried.</p><p>APIs are called.</p><p>Relevant memories are recalled.</p><p>Only after assembling all of that information does the language model begin reasoning.</p><p>This is why the smartest AI products aren’t necessarily powered by the smartest models.</p><p>They’re powered by the smartest retrieval systems.</p><p>The competitive advantage in modern AI is no longer just <strong>how well a model reasons.</strong></p><p>It’s <strong>how well a system finds the right information before reasoning even begins.</strong></p><p>In many ways, every successful AI application has quietly become something else.</p><p>A search engine.</p><p><strong>What We Think AI Looks Like</strong></p><p>The mental model is simple.</p><p>The user asks a question.</p><p>The model thinks.</p><p>The model answers.</p><p>Unfortunately, that’s almost never what happens in production.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1YRbA8mgJTqb6-GBKL0udg.png" /></figure><p><strong>The Hidden Architecture</strong></p><p>Modern AI systems don’t rely solely on what the model learned during training.</p><p>Instead, they retrieve the information they need when a question is asked.</p><p>Imagine asking your company’s internal AI assistant:</p><blockquote><em>“How many vacation days do contractors receive?”</em></blockquote><p>The model doesn’t magically know your company’s HR policy.</p><p>Instead, the application might:</p><ul><li>Search the HR knowledge base</li><li>Retrieve the latest policy document</li><li>Check whether you’re a contractor or a full-time employee</li><li>Load your previous conversation</li><li>Apply company-specific business rules</li><li>Assemble all of this into context</li><li>Finally, ask the LLM to generate an answer</li></ul><p>The model isn’t recalling facts from memory.</p><p>Its reasoning is based on information retrieved just seconds earlier.</p><p>That’s a fundamentally different way of building software.</p><p><strong>AI Doesn’t Know Everything</strong></p><p>One of the biggest misconceptions about large language models is that they function like enormous encyclopedias.</p><p>They don’t.</p><p>Foundation models possess broad world knowledge, but production AI products rarely depend on that knowledge alone.</p><p>Think about applications like:</p><ul><li>GitHub Copilot</li><li>Microsoft Copilot</li><li>ChatGPT Enterprise</li><li>Notion AI</li><li>Cursor</li><li>Perplexity</li></ul><p>None of these tools succeeds because they memorize your private data during training.</p><p>They succeed because they know <strong>where to look</strong>.</p><p>When you ask Cursor to explain a function, it searches your codebase.</p><p>When GitHub Copilot suggests code, it considers the files you’re editing.</p><p>When Perplexity answers a question, it retrieves fresh information from the web.</p><p>When ChatGPT Enterprise answers company-specific questions, it accesses enterprise documents you’ve connected.</p><p>The intelligence comes from combining retrieval with reasoning.</p><p>Not from reasoning alone.</p><p><strong>Search Has Quietly Become Infrastructure</strong></p><p>Traditional software applications were built around APIs.</p><p>Need customer information?</p><p>Call the Customer Service.</p><p>Need inventory?</p><p>Call the Inventory API.</p><p>Need payment status?</p><p>Call Stripe.</p><p>Everything revolved around structured requests.</p><p>Modern AI applications work differently.</p><p>Instead of asking:</p><blockquote><em>“Which API should I call?”</em></blockquote><p>They often ask:</p><blockquote><em>“What information should I retrieve before the model answers?”</em></blockquote><p>That subtle change has enormous architectural implications.</p><p>Retrieval is no longer an optional feature.</p><p>It’s becoming foundational infrastructure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uUvkQv_c5yHPDi6wOdGAeA.png" /></figure><p><strong>Search Isn’t Just Search Anymore</strong></p><p>When people hear the word <em>search</em>, they usually think about typing keywords into Google.</p><p>Modern AI retrieval is much more sophisticated.</p><p>Instead of matching exact words, AI systems search for meaning.</p><p>Ask these two questions:</p><blockquote><em>“How do I reset my password?”</em></blockquote><p>and</p><blockquote><em>“I forgot my login credentials.”</em></blockquote><p>A traditional keyword search might treat these as separate requests.</p><p>Modern retrieval systems recognize that they express the same intent.</p><p>This is semantic retrieval.</p><p>Instead of matching text, the system matches ideas.</p><p>That’s why AI applications can retrieve relevant documents even when users phrase questions in completely different ways.</p><p>The system isn’t searching for identical words.</p><p>It’s searching for a similar meaning.</p><p><strong>Retrieval Is Only the Beginning</strong></p><p>Finding documents is only half the challenge.</p><p>Imagine retrieving twenty highly relevant documents for a single question.</p><p>Should the model read all twenty?</p><p>Probably not.</p><p>Sending everything increases cost.</p><p>It increases latency.</p><p>It may even confuse the model.</p><p>The real engineering challenge is deciding <strong>which information the model should attend to</strong>.</p><p>This is where modern AI systems go beyond retrieval.</p><p>They perform something I like to call <strong>Context Assembly</strong>.</p><p><strong>Introducing Context Assembly</strong></p><p>Retrieval gives you information.</p><p>Context Assembly decides what the model should actually see.</p><p>This distinction is incredibly important.</p><p>A production AI system might retrieve:</p><ul><li>Three policy documents</li><li>Two previous conversations</li><li>A customer profile</li><li>Recent transaction history</li><li>Live inventory information</li><li>Business rules</li><li>External API responses</li></ul><p>Simply dumping everything into the prompt rarely produces the best answer.</p><p>Instead, the application filters, ranks, summarizes, compresses, and organizes the information before passing it to the model.</p><p>The goal isn’t to maximize context.</p><p>The goal is to maximize <strong>relevant context</strong>.</p><p>In other words, the system acts less like a search engine and more like an editor preparing research for an expert before they make a decision.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6QW-4B2kWhrXPXyMArxa6g.png" /></figure><p><strong>Why Context Quality Matters More Than Model Size</strong></p><p>Imagine two AI assistants.</p><p>The first uses the world’s most advanced language model but retrieves outdated, irrelevant documents.</p><p>The second uses a smaller model but consistently provides accurate, current, and highly relevant context.</p><p>Which one delivers better answers?</p><p>In practice, the second system often wins.</p><p>Because reasoning without information isn’t intelligence.</p><p>It’s guessing.</p><p>The model can only reason over what it sees.</p><p>If the context is incomplete, outdated, or irrelevant, even the most capable language model will struggle.</p><p>That’s why many production AI teams are discovering an unexpected truth:</p><blockquote><strong><em>Better retrieval often creates larger gains than switching to a bigger model.</em></strong></blockquote><p><strong>Retrieval Is Becoming the Competitive Advantage</strong></p><p>For years, AI companies competed by building better models.</p><p>Today, many companies use the same frontier models.</p><p>GPT, Claude. Gemini, Llama and Mistral.</p><p>The model itself is no longer the biggest differentiator.</p><p>The real competitive advantage lies in <strong>what happens before the prompt reaches the model.</strong></p><p>Imagine two companies using the exact same language model.</p><p>Company A simply sends the user’s question to the model.</p><p>Company B first:</p><ul><li>Retrieves relevant documentation.</li><li>Finds similar historical conversations.</li><li>Loads customer preferences.</li><li>Applies business rules.</li><li>Queries internal databases.</li><li>Reranks the retrieved information.</li><li>Removes redundant context.</li><li>Optimizes the prompt.</li></ul><p>Both companies use the same LLM.</p><p>Yet one consistently produces better answers.</p><p>The difference isn’t intelligence.</p><p>It’s information.</p><p>Modern AI products are beginning to compete on <strong>context quality</strong>, not just model quality.</p><p><strong>Better Retrieval Often Beats Bigger Models</strong></p><p>One of the most common assumptions in AI is that larger models automatically produce better products.</p><p>Reality is more nuanced.</p><p>Imagine asking an AI assistant:</p><blockquote><em>“Can I cancel my subscription after the free trial?”</em></blockquote><p>A powerful reasoning model won’t help if it retrieves last year’s pricing policy.</p><p>Now imagine a much smaller model that retrieves the latest policy, understands your subscription tier, and includes regional business rules.</p><p>Which answer is more useful?</p><p>The second one.</p><p>The smaller model succeeds because it starts with better information.</p><p>This changes how AI teams should think about optimization.</p><p>Instead of immediately upgrading to a larger model, ask:</p><ul><li>Is retrieval finding the right documents?</li><li>Is the information current?</li><li>Are we providing too much context?</li><li>Are we ranking results effectively?</li><li>Are we retrieving the right information for this specific user?</li></ul><p>Sometimes the biggest improvement comes from changing the retrieval layer — not the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iD-ery08iS0xh_cLeiI15w.png" /></figure><p><strong>Search Is No Longer Just Search</strong></p><p>Modern retrieval systems are much more sophisticated than simply finding documents. Most production AI systems combine multiple retrieval strategies.</p><p>Instead of relying on a single search mechanism, they blend several approaches to maximize relevance.</p><p>For example:</p><p><strong>Semantic Search</strong></p><p>Finds information based on meaning rather than exact keywords.</p><p><strong>Keyword Search</strong></p><p>Still valuable for product names, error codes, IDs, and technical terms.</p><p><strong>Metadata Filtering</strong></p><p>Limits results based on permissions, departments, locations, or timestamps.</p><p><strong>Reranking Models</strong></p><p>Re-evaluate retrieved documents to determine which are truly most relevant.</p><p><strong>Knowledge Graphs</strong></p><p>Connect entities and relationships that traditional search might overlook.</p><p>The best AI systems don’t rely on one retrieval technique.</p><p>They orchestrate many of them together.</p><p>The result isn’t just a better search.</p><p>It’s better reasoning.</p><p><strong>The Rise of Context Engineering</strong></p><p>Feature Engineering defined the last decade of machine learning.</p><p>Context Engineering is defining the next decade of AI.</p><p>Think about everything that can become context:</p><ul><li>Internal documentation</li><li>Previous conversations</li><li>User preferences</li><li>Calendar events</li><li>CRM records</li><li>SQL databases</li><li>PDFs</li><li>Images</li><li>Business policies</li><li>External APIs</li><li>Live web search</li><li>Organizational knowledge</li></ul><p>None of these pieces is useful on its own.</p><p>Their value lies in assembling them into the right context for a specific question.</p><p>That’s why building AI increasingly resembles building an information pipeline rather than simply deploying a model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fpzXyRjAFDl304cKK3ZTPw.png" /></figure><p><strong>Why Search Is Becoming Infrastructure</strong></p><p>The previous generation of software was built around databases.</p><p>Applications queried structured tables.</p><p>Business logic transformed the results.</p><p>Responses were deterministic.</p><p>Modern AI introduces an entirely new infrastructure layer.</p><p>Instead of asking databases directly, AI systems increasingly rely on retrieval platforms that combine structured and unstructured information into a unified context.</p><p>This new layer manages:</p><ul><li>Vector indexes</li><li>Search indexes</li><li>Embeddings</li><li>Metadata</li><li>Permissions</li><li>Freshness</li><li>Relevance</li><li>Memory</li><li>Context compression</li></ul><p>These aren’t product features.</p><p>They’re infrastructure.</p><p>Just as every web application eventually needed a database, every AI application is beginning to need a sophisticated retrieval layer.</p><p><strong>The Future AI Stack</strong></p><p>Over the next few years, I believe AI architectures will become increasingly standardized.</p><p>Instead of centering everything around the model, they’ll center around information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e6tQkaaHKCHsE3SFWJkJow.png" /></figure><p>Notice something interesting.</p><p>The LLM is no longer the center of the architecture.</p><p>It’s one component in a much larger system.</p><p>The intelligence emerges from how these components work together.</p><p><strong>The Shift From Models to Systems</strong></p><p>One of the biggest lessons of the past two years is that AI products aren’t defined by their models.</p><p>They’re defined by their systems.</p><p>A model can reason brilliantly.</p><p>But without relevant information, it reasons in the dark.</p><p>Conversely, even a modest model can produce remarkably useful responses when equipped with timely, relevant, and well-structured context.</p><p>That’s why engineering effort is shifting toward:</p><ul><li>Retrieval pipelines</li><li>Context assembly</li><li>Memory management</li><li>Evaluation frameworks</li><li>Agent orchestration</li><li>Tool integration</li><li>Observability</li><li>Continuous improvement</li></ul><p>The model is still essential.</p><p>But it’s no longer the entire product.</p><p><strong>Closing Thoughts</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7jSKhJufttNg_88K3iTBGA.png" /><figcaption>Intelligence Begins Before the First Token</figcaption></figure><p>When people interact with AI, they often focus on the answer.</p><p>Engineers know the real story begins much earlier.</p><p>Long before the model generates its first token, an invisible system is already making hundreds of decisions:</p><ul><li>What information matters?</li><li>What should be ignored?</li><li>Which documents are most relevant?</li><li>Which memories should be recalled?</li><li>Which tools should be used?</li><li>Which business rules must be enforced?</li></ul><p>These decisions shape the answer just as much as the language model itself.</p><p>That’s why I believe the future of AI won’t be won by the largest companies.</p><p>It will be won by companies that build the best information systems.</p><p>Because every great AI product is, at its core, a search engine that has learned how to reason.</p><p>The next frontier isn’t just artificial intelligence.</p><p>It’s <strong>information intelligence</strong>.</p><p><strong>What do you think?</strong></p><p>Do you believe <strong>retrieval and context engineering</strong> will become a bigger competitive advantage than larger foundation models over the next few years?</p><p>I’d love to hear your perspective in the comments — especially if you’re building production AI systems.</p><p>If you enjoyed this article, consider:</p><ul><li>👏 <strong>Clapping</strong> to help more engineers discover it.</li><li>💬 <strong>Sharing your thoughts</strong> or challenging the ideas in the comments.</li></ul><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d9QTaaaxboQP_gKSLedW_w.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://www.linkedin.com/in/bhuwan-kc-583505319/?isSelfProfile=true"><strong>Linkedin</strong></a>| <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p><strong>Disclosure:</strong> This post includes affiliate and partnership links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=227495efa4a4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/why-every-ai-product-is-secretly-a-search-engine-227495efa4a4">Why Every AI Product Is Secretly a Search Engine</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Cost Is Becoming an Accuracy Metric for any AI system]]></title>
            <link>https://ai.plainenglish.io/cost-is-becoming-an-accuracy-metric-72314f157ab6?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/72314f157ab6</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Fri, 03 Jul 2026 04:48:02 GMT</pubDate>
            <atom:updated>2026-07-09T03:45:01.855Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why the smartest AI system isn’t always the best one — and why engineering is becoming an economics problem</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rpj9uyo4zT7yLJGnrFSSlQ.png" /></figure><p>For most of machine learning&#39;s history, success is remarkably straightforward.</p><p>Build a model.</p><p>Measure its performance.</p><p>Choose the one with the highest accuracy.</p><p>If Model A achieved 95% accuracy and Model B achieved 92%, the decision is obvious. Deploy Model A.</p><p>Entire careers are built around squeezing out another percentage point of precision, recall, or F1 score. Feature engineering became an art. Hyperparameter tuning became a science. Competitions were won by fractions of a percent.</p><p>Accuracy is the ultimate objective.</p><p>Then generative AI changed the rules.</p><p>Today, two AI systems can produce nearly identical answers while having dramatically different costs, latencies, and business impact. One system may generate an answer for one cent in under a second, while another requires twelve cents and several seconds to answer the same question.</p><p>Both are technically “accurate.”</p><p>But are they equally good products?</p><p>Not necessarily.</p><p>The question modern AI teams increasingly ask isn’t:</p><blockquote><strong>“Which model is the most accurate?”</strong></blockquote><p>It’s:</p><blockquote><strong>“Which system creates the most value for every dollar we spend?”</strong></blockquote><p>That’s a fundamentally different optimization problem.</p><p><strong>The Era of Accuracy-First Machine Learning</strong></p><p>Traditional machine learning is largely an offline optimization exercise.</p><p>Everything revolved around improving predictive performance.</p><p>Teams tracked metrics like:</p><ul><li>Accuracy</li><li>Precision</li><li>Recall</li><li>ROC-AUC</li><li>F1 Score</li><li>Mean Absolute Error</li><li>RMSE</li></ul><p>When a new model improved accuracy — even slightly — it is often worth deploying.</p><p>Why?</p><p>Because inference is relatively cheap.</p><p>Whether a fraud detection model is scoring one customer or one million customers, the cost of each prediction is usually negligible compared to the value it creates.</p><p>Inference is mostly cloud cost.</p><p><strong>Generative AI Changed the Economics</strong></p><p>Large language models introduced something machine learning rarely had to worry about:</p><p><strong><em>Every prediction has a measurable cost.</em></strong></p><p>Every request consumes:</p><ul><li>Compute</li><li>GPU time</li><li>Tokens</li><li>Network bandwidth</li><li>Retrieval operations</li><li>Tool executions</li><li>API calls</li></ul><p>Unlike traditional ML, intelligence is no longer a one-time training expense.</p><p>It’s an ongoing operational cost.</p><p>Imagine building a customer support assistant.</p><p>Every customer interaction might involve:</p><ul><li>Retrieving documents from a vector database</li><li>Searching product documentation</li><li>Calling internal APIs</li><li>Running multiple prompts</li><li>Using an evaluator model</li><li>Generating a final response</li></ul><p>A single user question might invoke five different AI components before an answer is produced.</p><p>Each component has a cost.</p><p>Individually, they seem insignificant.</p><p>At enterprise scale, they’re enormous.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F8fu1U2mYbq5n1_J29ZQ-g.png" /></figure><p><strong>Every Token Has a Price</strong></p><p>One of the biggest mindset shifts for engineers moving into AI is realizing that every token matters.</p><p>Longer prompts increase cost.</p><p>Longer responses increase cost.</p><p>Repeated retrieval increases cost.</p><p>Multiple tool calls increase cost.</p><p>Reflection and self-correction increase cost.</p><p>Even asking an LLM to “think harder” often means spending more money.</p><p>Unlike classical software, where executing additional lines of code was practically free, modern AI systems pay for intelligence one request at a time.</p><p>The smarter you ask the system to be, the more expensive it becomes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eXV6eW_QCNxfOiE9hsWG2Q.png" /></figure><p><strong>The New Optimization Equation</strong></p><p>Machine learning optimized for prediction quality.</p><p>AI products optimize for business value.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*26EsAWIy7ZZR9trdUm5Egg.png" /></figure><p>Improving any one of these variables affects the overall value delivered by the system.</p><p>An answer that is marginally more accurate but dramatically slower and more expensive may reduce the product&#39;s overall effectiveness.</p><p>Success is no longer defined by one metric.</p><p>It’s defined by balance.</p><p><strong>When the Best Model Is the Wrong Choice</strong></p><p>Consider two AI assistants.</p><p><strong>Assistant A</strong></p><ul><li>97% response quality</li><li>8 seconds average latency</li><li>$0.12 per request</li></ul><p><strong>Assistant B</strong></p><ul><li>95% response quality</li><li>1.5 seconds latency</li><li>$0.01 per request</li></ul><p>At first glance, Assistant A appears superior.</p><p>But imagine your application serves 200 million requests every year.</p><p>The additional cost becomes millions of dollars.</p><p>Meanwhile, most users may never notice the 2% difference in quality.</p><p>They will notice waiting eight seconds for every answer.</p><p>The “best” model on paper may produce the worst user experience.</p><p><strong>Cost Is No Longer Just a Finance Problem</strong></p><p>Historically, infrastructure teams worried about cloud bills.</p><p>ML engineers worried about model accuracy.</p><p>Those responsibilities are beginning to merge.</p><p>Modern AI engineers constantly make architectural decisions that influence cost.</p><p>Should this request use GPT-5?</p><p>Or is a smaller model sufficient?</p><p>Should we retrieve five documents?</p><p>Or would two high-quality documents be enough?</p><p>Should we summarize the conversation?</p><p>Or send the entire history?</p><p>Should we invoke another tool?</p><p>Or answer immediately?</p><p>Every architectural decision now has financial consequences.</p><p>Engineering has become economics.</p><p><strong>Bigger Models Aren’t Always Smarter Products</strong></p><p>A common misconception in AI is that larger models automatically produce better systems.</p><p>In reality, larger models often solve problems that didn’t exist.</p><p>Imagine asking an AI assistant:</p><blockquote><strong>“What’s your refund policy?”</strong></blockquote><p>Does this require a frontier reasoning model?</p><p>Probably not.</p><p>A lightweight model, combined with effective retrieval, can answer perfectly.</p><p>Now imagine asking:</p><blockquote><strong>“Compare three acquisition strategies for expanding into Southeast Asia while considering regulatory constraints and projected currency fluctuations.”</strong></blockquote><p>That’s a different problem entirely.</p><p>Not every request deserves the most powerful model available.</p><p>The smartest systems don’t rely on a single model.</p><p>They use the right model at the right time.</p><p><strong>The Rise of AI Routing</strong></p><p>One of the most important architectural patterns emerging in production AI is intelligent routing.</p><p>Instead of sending every request to the same model, systems decide where each request should go.</p><p>Simple questions may use a small, inexpensive model.</p><p>Moderately complex questions might use a mid-sized model.</p><p>Only the most challenging reasoning tasks reach the largest model.</p><p>This approach improves:</p><ul><li>Cost</li><li>Latency</li><li>Scalability</li></ul><p>without sacrificing user experience.</p><p>In many cases, routing creates more business value than upgrading to a larger model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xKqUPqJZWb10cPHurokVIg.png" /></figure><p><strong>Retrieval Is Often Cheaper Than Reasoning</strong></p><p>Many AI teams instinctively solve poor responses by switching to a larger model.</p><p>Often, the real problem isn’t intelligence.</p><p>It’s missing information.</p><p>A better retrieval pipeline can outperform a larger model because it provides the right context, rather than expecting the model to infer everything.</p><p>This is why investments in retrieval, reranking, and context engineering frequently produce larger gains than simply increasing model size.</p><p>Sometimes the cheapest improvement is also the most effective.</p><p><strong>Caching Is an Intelligence Multiplier</strong></p><p>Traditional software engineers have used caching for decades to improve performance.</p><p>AI systems are rediscovering their value.</p><p>If thousands of users ask the same question every day, why generate the answer from scratch each time?</p><p>Semantic caching allows systems to reuse high-quality responses for similar requests.</p><p>The result?</p><ul><li>Lower latency</li><li>Lower cost</li><li>Higher consistency</li></ul><p>Caching doesn’t make the model smarter.</p><p>It makes the system more efficient.</p><p><strong>Measuring ROI Instead of Accuracy</strong></p><p>As AI becomes embedded in products, executives increasingly ask different questions.</p><p>Instead of:</p><blockquote><strong>“What’s the model’s accuracy?”</strong></blockquote><p>They ask:</p><ul><li>How much does each response cost?</li><li>How many users does it retain?</li><li>Does it reduce support tickets?</li><li>Does it increase productivity?</li><li>What’s the return on investment?</li></ul><p>These are product questions.</p><p>The AI model is only one contributor to the answer.</p><p><strong>The Future Is Cost-Aware AI</strong></p><p>Future AI systems will optimize dynamically.</p><p>They’ll decide:</p><ul><li>Which model to use?</li><li>Whether retrieval is necessary.</li><li>When to call external tools.</li><li>Whether to summarize or compress context.</li><li>Whether a cached answer is sufficient.</li><li>When a human should be involved.</li></ul><p>These decisions won’t just improve efficiency.</p><p>They’ll become a competitive advantage.</p><p>Companies that intelligently allocate intelligence will outperform companies that simply buy more compute.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JPz6NpKhElXGKYz2mulVzA.png" /><figcaption>Great AI products don’t maximize quality. They optimize the balance between quality and cost.</figcaption></figure><p><strong>Closing Thoughts</strong></p><p>For years, machine learning rewarded teams that optimized for model accuracy.</p><p>Modern AI rewards teams that optimize for outcomes.</p><p>The most successful AI systems won’t necessarily have the largest models or the highest benchmark scores.</p><p>They’ll deliver the right answer, at the right time, using the right amount of intelligence.</p><p>Because in production AI, intelligence isn’t free.</p><p>It’s a resource that must be managed wisely.</p><p>As engineers, we’re no longer optimizing only for correctness.</p><p>We’re optimizing for <strong>quality, reliability, latency, scalability, and cost</strong> — all at once.</p><p>In that world, cost doesn’t compete with accuracy.</p><p>It becomes part of how we define it.</p><p>The next generation of AI engineering won’t be measured by <strong>who builds the smartest model</strong>.</p><p>It will be measured by <strong>who builds the smartest system</strong>.</p><p>And the smartest systems will understand a lesson that traditional machine learning rarely had to learn:</p><blockquote><strong>The best answer isn’t always the most intelligent one. It’s the one that creates the greatest value.</strong></blockquote><p><strong>Do you agree, or do you think model quality will remain the biggest competitive advantage?</strong></p><p>I’d genuinely love to hear your perspective in the comments.</p><p>If you’re interested in thoughtful discussions on AI engineering and production AI, follow along — there’s much more to explore.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72314f157ab6" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/cost-is-becoming-an-accuracy-metric-72314f157ab6">Cost Is Becoming an Accuracy Metric for any AI system</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Human Feedback Is Becoming a First-Class Dataset in AI systems]]></title>
            <link>https://ai.plainenglish.io/human-feedback-is-becoming-a-first-class-dataset-11f8a7439903?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/11f8a7439903</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Thu, 02 Jul 2026 01:16:00 GMT</pubDate>
            <atom:updated>2026-08-15T15:03:14.855Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Why the most valuable data in AI doesn’t exist until users start using your product.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DDAL8lDEh1nzednXqqt9Yw.png" /></figure><p>The core insight is:</p><blockquote><strong><em>In traditional ML, datasets were created before deployment. In modern AI, some of the most valuable datasets are created after deployment.</em></strong></blockquote><p>That’s a profound change.</p><p><strong>Hook</strong></p><p>For decades, machine learning began with a dataset. Engineers collected data, cleaned it, labeled it, trained a model, and deployed it. Deployment marked the end of the data collection process. Modern AI has turned that workflow upside down.</p><p>Today, deployment is often the beginning of your most valuable dataset. Every thumbs up, correction, regenerated response, edit, clarification, abandoned conversation, and follow-up question is data. Human feedback is no longer just a UX metric. It’s becoming a first-class dataset.</p><p><strong>The Traditional ML Mindset vs The New AI Feedback Loop</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BHe5tcFtSOZHUHOY2x21vA.png" /></figure><p><strong>AI Products Never Stop Learning</strong></p><p>An AI application generates new information every day.</p><p>For example:</p><ul><li>Users rewrite generated emails.</li><li>Developers edit generated code.</li><li>Customers regenerate responses.</li><li>Users click “thumbs down.”</li><li>They ask follow-up questions.</li><li>They abandon conversations midway.</li><li>They choose one suggestion over another.</li></ul><p>These aren’t just interactions.</p><p>They’re labels created organically by real users.</p><p><strong>Feedback Is the New Labeling Process</strong></p><p>Traditional supervised learning relied on expensive annotation teams. Modern AI products often receive continuous user feedback without explicitly asking for it.</p><p>Examples include:</p><blockquote><em>👍 </em><strong><em>Thumbs Up</em></strong><em> → The response was useful.</em></blockquote><blockquote><em>👎 </em><strong><em>Thumbs Down</em></strong><em> → The response failed.</em></blockquote><blockquote><em>🔄 </em><strong><em>Regenerate</em></strong><em> → The user expected a better answer.</em></blockquote><blockquote><em>📋 </em><strong><em>Copy Response</em></strong><em> → The response had high utility.</em></blockquote><blockquote><em>✏️ </em><strong><em>Edit Response</em></strong><em> → The response was partially correct.</em></blockquote><blockquote><em>🚫 </em><strong><em>Ignore Response</em></strong><em> → The response wasn’t relevant.</em></blockquote><blockquote><em>❓ </em><strong><em>Follow-up Question</em></strong><em> → The model lacked sufficient context.</em></blockquote><blockquote><em>💬 </em><strong><em>Long Conversation</em></strong><em> → The interaction was engaging.</em></blockquote><p>Every interaction teaches the system something.</p><p><strong>Why This Data Is More Valuable</strong></p><p>Human feedback has properties that traditional datasets often lack:</p><ul><li>Real-world</li><li>Task-specific</li><li>Continuously updated</li><li>Reflects actual user expectations</li><li>Captures changing preferences</li></ul><p>Instead of asking annotators,</p><blockquote><em>“Is this response correct?”</em></blockquote><p>You’re observing whether real users found it helpful.</p><p>That signal is often richer.</p><p><strong>Not All Feedback Should Retrain the Model</strong></p><p>One of the biggest misconceptions is that every piece of feedback should be used for model fine-tuning. In practice, feedback can improve systems in many ways:</p><ul><li>Better prompts</li><li>Improved retrieval</li><li>Updated business rules</li><li>Personalization</li><li>Context selection</li><li>Tool routing</li><li>Agent planning</li><li>Future evaluations</li></ul><p>Often, improving the system around the model delivers more value than retraining the model itself.</p><p><strong>AI Teams Need Feedback Engineering</strong></p><p>Collecting feedback isn’t enough. You need to answer questions like:</p><ul><li>Which feedback is trustworthy?</li><li>How do we aggregate conflicting signals?</li><li>How do we detect noisy users?</li><li>Which interactions deserve human review?</li><li>When should feedback update memory?</li><li>When should it trigger retraining?</li></ul><p>This is an engineering problem, not just a product problem.</p><p><strong>The Biggest Competitive Advantage</strong></p><p>As foundation models become increasingly similar, the differentiator shifts. Every AI company can access powerful models. What they can’t copy is the proprietary feedback generated by their users. That feedback becomes a unique asset. Over time, it shapes:</p><ul><li>Better evaluations</li><li>Better retrieval</li><li>Better prompts</li><li>Better workflows</li><li>Better personalization</li></ul><p>The product improves because its users teach it.</p><p><strong>Closing Thoughts</strong></p><p>For years, data science has focused on collecting better training data before deployment. AI engineering is increasingly focused on collecting better feedback after deployment. The dataset is no longer frozen during training. It’s alive. And every user interaction becomes an opportunity to make the system smarter. The future of AI won’t belong solely to the teams with the largest models. It will belong to the teams that learn most effectively from their users.</p><p><em>What are your thoughts? Looking forward to hearing the different perspective.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=11f8a7439903" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/human-feedback-is-becoming-a-first-class-dataset-11f8a7439903">Human Feedback Is Becoming a First-Class Dataset in AI systems</a> was originally published in <a href="https://ai.plainenglish.io/">Artificial Intelligence in Plain English</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Context Is the New Feature Engineering]]></title>
            <link>https://medium.com/codetodeploy/context-is-the-new-feature-engineering-e27885e1013f?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/e27885e1013f</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Wed, 01 Jul 2026 04:03:07 GMT</pubDate>
            <atom:updated>2026-08-03T11:22:57.591Z</atom:updated>
            <content:encoded><![CDATA[<p>The competitive advantage in AI is no longer just the model you choose — it’s the context you provide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XorqkdcYVNToUfsk3PmCmQ.png" /></figure><p><strong>Hook</strong></p><p>Imagine asking two engineers to build the same AI assistant.</p><blockquote><strong><em>Master the Data &amp; AI Skills Top Employers Are Hiring For.</em></strong><br><em> Get career-ready with hands-on, practical training.</em><br> <em>🔥 50% OFF — Limited-Time Offer!</em><br><a href="https://datacamp.pxf.io/gR4Rv0"> 👉 <strong>Enroll Now &amp; Start Learning</strong></a></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pkQFf0NUnzoFbVEVQCF5WQ.png" /></figure><p>They both use the same model.<br>They both use the same API.<br>They both spend the same amount on inference.</p><p>One assistant feels remarkably intelligent. The other hallucinates, misses obvious details, and gives generic responses.</p><p>The difference isn’t the model. It’s the context.</p><p>For years, machine learning engineers competed by building better features. Today, AI engineers compete by building better context.</p><p>The hardest part of modern AI systems is no longer training models. It’s engineering the information that those models see before they generate an answer.</p><p><strong>The Age of Feature Engineering</strong></p><p>Traditional machine learning revolved around one question:</p><blockquote><em>How can we transform raw data into something the model can understand?</em></blockquote><p>A fraud detection model didn’t learn directly from transaction logs. Engineers created features like:</p><ul><li>Average transaction amount</li><li>Number of purchases in the last 24 hours</li><li>Geographic distance between purchases</li><li>Device consistency</li><li>Historical customer behavior</li></ul><p>The model’s performance often depended more on feature quality than on the choice of algorithm.</p><p>Feature engineering became one of the defining skills of every data scientist.</p><p><strong>Then Foundation Models Changed Everything</strong></p><p>Large language models arrived with a surprising property.</p><p>They already understood language.</p><p>They already knew programming.</p><p>They already possessed a vast amount of world knowledge.</p><p>Suddenly, engineers weren’t asking:</p><blockquote><em>“What features should I build?”</em></blockquote><p>Instead, they asked:</p><blockquote><em>“What information should the model see before answering?”</em></blockquote><p>That shift fundamentally changed AI engineering.</p><p><strong>Context Engineering Has Replaced Feature Engineering</strong></p><p>Modern AI systems rarely fail because the model lacks intelligence.</p><p>They fail because the model lacks context.</p><p>Context can include:</p><ul><li>User history</li><li>Company documents</li><li>Database records</li><li>API responses</li><li>Previous conversations</li><li>Images</li><li>Business rules</li><li>Tool outputs</li><li>Memory</li><li>Real-time data</li></ul><p>Every successful AI application is really a context assembly system.</p><p>The LLM is simply the reasoning engine at the end.</p><p><strong>Context Is More Than Retrieval</strong></p><p>Many people equate context engineering with RAG.</p><p>That’s only one piece.</p><p>Good context engineering involves deciding:</p><ul><li>What information should be retrieved?</li><li>How much should be retrieved?</li><li>In what order?</li><li>What should be summarized?</li><li>What should be omitted?</li><li>What should be structured instead of copied?</li><li>What should be refreshed in real time?</li></ul><p>These decisions often have a larger impact than switching to a newer model.</p><p><strong>Context Has Become the New Data Pipeline</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1rXSUl8btrktOijZC4IwyA.png" /></figure><p>The pipeline hasn’t disappeared — it’s just moved from data engineering into context engineering. Same shape, different inputs at every stage.</p><p><strong>Bigger Context Doesn’t Mean Better Context</strong></p><p>A common misconception is that larger context windows solve everything.</p><p>They don’t.</p><p>An LLM with 1 million tokens of irrelevant information can perform worse than one with 5,000 carefully selected tokens.</p><p>Quality beats quantity.</p><p>The challenge isn’t maximizing context.</p><p>It’s maximizing relevant context.</p><p><strong>The New Responsibilities of AI Engineers</strong></p><p>The role of the AI engineer increasingly resembles that of an information architect.</p><p>Instead of creating features, they design systems that determine:</p><ul><li>What the model should know</li><li>When should it know it</li><li>Where that information comes from</li><li>How trustworthy is it</li><li>How long does it remain valid</li></ul><p>In many ways, AI engineers are becoming context engineers.</p><p><strong>Why This Matters</strong></p><p>As foundation models continue to improve, the performance gap between them is shrinking.</p><p>The difference between great AI products and mediocre ones increasingly comes down to questions like:</p><ul><li>Did the model retrieve the right document?</li><li>Was the conversation history relevant?</li><li>Were business rules included?</li><li>Did memory improve personalization?</li><li>Was the external data fresh?</li></ul><p>These aren’t model problems.</p><p>They’re context problems.</p><p>And context is quickly becoming the primary competitive advantage in modern AI systems.</p><p><strong>Closing Thoughts</strong></p><p>For a decade, machine learning rewarded engineers who built better features.</p><p>The next decade will reward engineers who build better context.</p><p>Models will continue to become cheaper, faster, and more capable.</p><p>But no model can reason about information it never receives.</p><p>The future of AI isn’t just about larger models.</p><p>It’s about a smarter context.</p><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d9QTaaaxboQP_gKSLedW_w.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://www.linkedin.com/in/bhuwan-kc-583505319/?isSelfProfile=true"><strong>Linkedin</strong></a>| <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p><strong>Disclosure:</strong> This post includes affiliate and partnership links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e27885e1013f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/context-is-the-new-feature-engineering-e27885e1013f">Context Is the New Feature Engineering</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evaluation Is the New Training]]></title>
            <link>https://levelup.gitconnected.com/evaluation-is-the-new-training-1328f73b0c87?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/1328f73b0c87</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[llm-applications]]></category>
            <category><![CDATA[ai-systems]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Tue, 30 Jun 2026 09:11:00 GMT</pubDate>
            <atom:updated>2026-07-08T16:29:12.483Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B7DrZqU96mEzYd-DVBQI4w.png" /></figure><p>The hardest part of building AI systems has quietly moved — and most teams haven’t noticed.</p><p>For the first decade of applied machine learning, the game was simple to describe, even when it was hard to win: get labeled data, define a metric, train a model until the number goes up. Evaluation was almost administrative — compute accuracy, F1, AUC, and ship a report. The hard part was the model. Evaluation was the finish line.</p><p>That has flipped. And the inversion is more disorienting than most people admit.</p><p><strong>What the Old World Looked Like</strong></p><p>In classical supervised ML, evaluation was straightforward because the problem itself was structured. You had a label — a ground truth that someone had attached to each data point — and the entire system existed to predict it. “Is this transaction fraudulent?” has a right answer. “Will this customer churn?” has a right answer. You might argue over which metric to optimize or how to handle class imbalance, but the problem&#39;s shape was clear.</p><p>This meant evaluation had a clean contract: split your data, run inference, compare predictions to labels, and report a number. Do it rigorously, and you know where you stand. The hard problems were on the training side — data quality, feature engineering, hyperparameter tuning search, compute, and generalization. Evaluation surfaced those problems, but wasn’t itself a hard problem.</p><p><strong>What Changed</strong></p><p>The bottleneck moved when building became easy. Spinning up an LLM-powered system today takes hours, not months. You can have a working prototype of a document Q&amp;A tool, a customer support agent, or a code reviewer by the end of the afternoon. The API handles the model. Prompting handles the “feature engineering.” Infrastructure is largely commoditized.</p><p>This is genuinely remarkable — but it created a new gap that the industry hasn’t fully reckoned with.</p><p>If you ask most teams building LLM applications how they know their system is working, you get variations of the same answer: they try it. Someone on the team has some queries, sees if the responses look right, and flags obvious failures. This is sometimes called a “vibe check” — and it’s the dominant evaluation methodology in production AI systems right now.</p><p>The vibe check is not nothing. Human intuition catches real problems. But it doesn’t scale, it isn’t reproducible, and it has a dangerous blind spot: it catches the failures that look like failures. It misses the ones that look like success.</p><p><strong>The Core Problem</strong></p><p>Classical ML evaluation worked because the output space was small and the correctness criterion was external. The output a 0 or a 1. An independent source of truth says what the right answer was. The two are compared.</p><p>LLM outputs break both of these assumptions simultaneously.</p><p>The output space is enormous — a response to a question can be correct, partially correct, technically correct but misleading, hallucinated with high confidence, or correct in content but wrong in tone for the use case. There is no canonical “right answer” to compare against. And the correctness criterion is often internal to the response itself: you have to read and understand the output to evaluate it, which means evaluation requires nearly as much intelligence as generation.</p><p>For agents, this problem deepens significantly. A multi-step agent doesn’t just produce one output — it produces a trajectory of decisions, tool calls, and intermediate states before arriving at a final answer. Any step in that chain can fail silently while the downstream output looks plausible. An agent that retrieves the wrong document, reasons incorrectly about it, and then produces a confident-sounding response has failed in a way that a single output evaluation will completely miss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZtBnmQYnO90tLzc3jzHGkA.jpeg" /></figure><p><strong>What Teams Are Actually Doing — and Where It Breaks</strong></p><p>A few evaluation patterns have emerged as defaults, each with real limits:</p><p><strong>Golden set testing</strong> — a curated set of inputs with expected outputs that you run before every deployment. This works until the system changes enough that your golden set goes stale, or until your golden set is small enough that passing it tells you almost nothing about real-world performance. Most golden sets are both.</p><p><strong>LLM-as-judge </strong>— using a second language model to score the output of the first. This is genuinely useful, especially for measuring coherence, relevance, and style at scale. But LLM judges have their own biases — they prefer longer responses, they’re influenced by format, and they have blind spots that correlate with the blind spots of the model being evaluated. Using an LLM to evaluate an LLM is like asking someone to grade their own exam.</p><p><strong>User feedback signals </strong>— thumbs up/down, escalation rates, and session abandonment. These are real signals, but they’re lagged, sparse, and indirect. By the time user feedback surfaces a problem, many users have already experienced it.</p><p>None of these is wrong. The problem is treating any one of them as sufficient when evaluation now requires a portfolio approach, as training once did.</p><p><strong>The Shape of the Solution</strong></p><p>What a good evaluation infrastructure actually looks like in this new world is becoming clearer, even if most teams aren’t there yet.</p><p>It’s continuous rather than a pre-deployment checkpoint. The system is being evaluated in production all the time, not just before a release — behavioral monitoring that watches the distribution of outputs, flags anomalies, and tracks metrics that matter to the use case rather than generic quality scores.</p><p>It’s behavioral rather than output-level. Instead of asking “is this response good?”, it asks “does the system behave correctly across the full range of situations it will encounter?” — including adversarial inputs, edge cases, and the long tail of user queries that never appear in your test set.</p><p>It’s multi-layered. Component-level evals for individual tools and retrieval steps, trajectory evals for multi-step agents, end-to-end evals for user outcomes, and red-teaming to actively break the system. These answer different questions, and none of them can substitute for the others.</p><p>And critically, it treats evaluation as a product, not a task. Teams that do this well have people whose job is to build evaluation tooling — frameworks, datasets, judges, dashboards — not just to run evals before a release.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K7tVNZ9bOtgJ9iKreEd_9Q.jpeg" /></figure><p><strong>What This Means for the People Building These Systems</strong></p><p>The mental model shift is this: in classical ML, you built a model and then checked if it worked. In the LLM/agent era, you build a system and then build a second system to tell you if the first one is working — and that second system is often harder to get right.</p><p>This is uncomfortable because the first system is the one that looks impressive. Demos are made of outputs, not evaluation harnesses. But the teams that are actually shipping reliable AI products are the ones who’ve figured out that the evaluation infrastructure is the hard part, and invested accordingly.</p><p>The question worth sitting with isn’t “does this work?” — it’s “how would I know if it stopped working tomorrow, and would I know before my users did?”</p><p>That’s the question classical ML evaluation was never designed to answer. And answering it is where the real work is now.</p><p><em>Where have you landed on this? Are you running systematic evals, or still relying on the vibe check? I’d be curious what’s actually working in practice.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1328f73b0c87" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/evaluation-is-the-new-training-1328f73b0c87">Evaluation Is the New Training</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Software Engineering Is Becoming Probabilistic (And Why That’s Uncomfortable)]]></title>
            <link>https://medium.com/codetodeploy/why-software-engineering-is-becoming-probabilistic-and-why-thats-uncomfortable-1085f67c40a5?source=rss-f9dfd297d92------2</link>
            <guid isPermaLink="false">https://medium.com/p/1085f67c40a5</guid>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Akshay Adlakha]]></dc:creator>
            <pubDate>Tue, 30 Jun 2026 01:02:37 GMT</pubDate>
            <atom:updated>2026-08-16T05:35:35.172Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hUguS4nv4heqvmG9aBrRZw.png" /></figure><p>For most of the history of software engineering, we have operated on a single comforting assumption: same input, same output, every time. A function takes 1 and 2, and it returns 3. That assumption lets us build everything from accounting systems to flight control software on a foundation of predictability. If something broke, it broke the same way twice — which meant we could actually find and fix it.</p><blockquote><em>💥 </em><strong><em>Master AI &amp; Tech Skills </em></strong><br> 💥 Crack Every Tech Interview with Confidence<br> 🔥 Up to <strong>70% OFF</strong> — Limited-Time Offer<br> <a href="https://www.educative.io/unlimited?aff=xkRD"><em>👉 </em><strong><em>Enroll Now &amp; Start Learning</em></strong></a></blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ITogvtC6mF2BUF35Qsd9pg.png" /></figure><p>That foundation is shifting, and it’s worth being precise about why, instead of just waving at “AI is changing everything”.</p><p><strong>Two Different Contracts With Reality</strong></p><p>A deterministic system is a contract: given X, you get Y, no negotiation. If-this-then-that, all the way down. This is what compilers, databases, and traditional business logic have given us for decades, and it’s why we trust them with money, health records, and safety-critical infrastructure.</p><p>A probabilistic system makes a different kind of promise: given X, you will get something likely to be useful, with a distribution of possible outputs rather than one fixed answer. Machine Learning models, LLMs, and agentic systems all live here. Ask the same question twice, and you may get two different — both reasonable — answers.</p><p>Until recently, most production software kept two worlds carefully separated. ML models lived in their own services, behind APIs, producing scores and predictions that deterministic code would then act on with fixed rules. The “intelligence” was contained; the system around it stayed predictable.</p><p>That separation is breaking down, specifically because of agents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2E5tO6Bz9celL1TCbGDdGQ.jpeg" /></figure><p><strong>What Changed: From Predictions to Decisions</strong></p><p>A model that outputs a probability is easy to keep in a box. A model that decides which API to call, which file to edit, or which step to take next is a different animal entirely — it’s not predicting anymore, it’s acting. And once a system is making sequences of decisions rather than single predictions, the old engineering posture — “I own every path the data can take” — no longer holds.</p><p>Think of it as the difference between being a traffic controller and being a dispatcher. A traffic controller owns the intersection: every signal, every lane, every possible path is theirs to define. A dispatcher gives a driver a destination and general instructions, and then has to live with the fact that the driver might take a shortcut, hit traffic, or make a judgment call the dispatcher didn’t anticipate. Agentic systems put engineers in the dispatcher’s seat, often for the first time in their careers.</p><p>This is comfortable for a simple reason: most of us were implicitly or explicitly trained to treat consistency as the goal. A bug that happens the same way every time is a bug you can fix. A bug that happens sometimes, depending on phrasing, context window, or random sampling, is a different kind of problem.</p><p><strong>It’s Not Actually New — We Just Didn’t Admit It</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rhYo7D2zmWNUMgBIEqgFDA.jpeg" /></figure><p>Here’s the part that’s easy to miss: large-scale systems have quietly been probabilistic for a long time. Cache hits and misses, network retries, distributed consensus under partial failure, even floating-point rounding in certain contexts — engineers have been managing statistical behavior under a deterministic-looking surface for years. We just built enough abstraction layers that probabilistic parts stayed invisible to the average developer.</p><p>What’s different now isn’t that probabilism arrived — it’s that it moved from the infrastructure layer, where we had already built decades of tooling to name it, to the decision layer, where we have almost none.</p><p><strong>So What Does Responsible Engineering Look Like Here?</strong></p><p>The honest answer is: not “agentic everything,” and not “reject it and stay deterministic.” The pattern that’s actually working in production systems is closer to architectural discipline — deciding deliberately where each mode belongs, rather than reaching for the same tool everywhere.</p><p>A few principles seem to be holding up:</p><blockquote><strong>Scope the blast radius before you scope the intelligence.</strong> Let a probabilistic component propose, and keep a deterministic boundary around what it’s allowed to actually execute — structure outputs, validated inputs, explicit approval gates for anything irreversible.</blockquote><blockquote><strong>Keep the deterministic core where the stakes are high.</strong> Financial transactions, compliance logic, and safety-critical paths still want fixed rules. Probabilistic components are augmentation, not replacement, for these.</blockquote><blockquote><strong>Build observability for distributions, not just failures.</strong> A deterministic system tells you when it’s broken. A probabilistic one can be quietly “working” while drifting — you need to watch the shape of its outputs over time, not just whether it crashed.</blockquote><blockquote><strong>Treat evaluation as a discipline of its own. </strong>Testing a deterministic function is often enough. Testing a probabilistic component means sampling, statistical confidence, and accepting that “passing” is now a matter of degree.</blockquote><p><strong>The Real Shift</strong></p><p>The move toward probabilistic software isn’t really about LLMs being trendy. It’s that we’re handling more and more decisions — not just predictions — to systems that don’t guarantee a single right answer. That requires a different relationship with control: less “I specified every path” and more “I bounded the space of acceptable outcomes and built the guardrails to catch the rest.”</p><p>Engineers who are good at this aren’t the ones who’ve memorized prompt patterns. They’re the ones who can look at a system and correctly answer one question for every component: does this need a guarantee, a good-enough answer, or a fast answer? Getting that judgment right — component by component — is quickly becoming one of the more valuable skills in the field.</p><p>What’s your experience been — where have you drawn the deterministic/probabilistic line in systems you’ve built, and has it held up?</p><h3>Thank you for being a part of the community</h3><p><em>Before you go:</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d9QTaaaxboQP_gKSLedW_w.png" /></figure><p>👉 Be sure to <strong>clap</strong> and <strong>follow</strong> the writer ️👏<strong>️️</strong></p><p>👉 Follow us: <a href="https://www.linkedin.com/in/bhuwan-kc-583505319/?isSelfProfile=true"><strong>Linkedin</strong></a>| <a href="https://medium.com/codetodeploy"><strong>Medium</strong></a></p><p>👉 CodeToDeploy Tech Community is live on Discord — <a href="https://discord.gg/ZpwhHq6D"><strong>Join now!</strong></a></p><p><strong>Disclosure:</strong> This post includes affiliate and partnership links.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1085f67c40a5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/codetodeploy/why-software-engineering-is-becoming-probabilistic-and-why-thats-uncomfortable-1085f67c40a5">Why Software Engineering Is Becoming Probabilistic (And Why That’s Uncomfortable)</a> was originally published in <a href="https://medium.com/codetodeploy">CodeToDeploy</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>