<?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 Anagha Mulloth on Medium]]></title>
        <description><![CDATA[Stories by Anagha Mulloth on Medium]]></description>
        <link>https://medium.com/@anagharamdas2000?source=rss-a33659c77ad9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*htPrCxg9Zq6OzHbN3y7MmA.jpeg</url>
            <title>Stories by Anagha Mulloth on Medium</title>
            <link>https://medium.com/@anagharamdas2000?source=rss-a33659c77ad9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 18:41:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anagharamdas2000/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[Knowledge Graphs — The Only Guide You Need in 2026]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/knowledge-graphs-the-only-guide-you-need-in-2026-36dc3e78984a?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/1024/1*g60zMhReptHfrVsdLi1b7w.jpeg" width="1024"></a></p><p class="medium-feed-snippet">Generative AI has completely reshaped the tech landscape over the last few years. But if you have been working behind the scenes in 2026&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/knowledge-graphs-the-only-guide-you-need-in-2026-36dc3e78984a?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/knowledge-graphs-the-only-guide-you-need-in-2026-36dc3e78984a?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/36dc3e78984a</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 04:31:03 GMT</pubDate>
            <atom:updated>2026-10-21T18:27:40.467Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[RAG Debugging: Five Places Your Answer Broke Before the Model Ever Saw It]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/rag-debugging-five-places-your-answer-broke-before-the-model-ever-saw-it-7bcb850b0ab3?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/2600/1*fKkq6ZPbQTcqCAbPwJbXPw.jpeg" width="2816"></a></p><p class="medium-feed-snippet">You ask your system something about your own documents. The answer comes back wrong. Not comically wrong, which you&#x2019;d catch instantly, but&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/rag-debugging-five-places-your-answer-broke-before-the-model-ever-saw-it-7bcb850b0ab3?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/rag-debugging-five-places-your-answer-broke-before-the-model-ever-saw-it-7bcb850b0ab3?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/7bcb850b0ab3</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[rags]]></category>
            <category><![CDATA[generative-ai-solution]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Wed, 23 Sep 2026 11:55:58 GMT</pubDate>
            <atom:updated>2026-09-23T11:55:58.868Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Why Your RAG Pipeline Works in Testing but Fails in Production]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/why-your-rag-pipeline-works-in-testing-but-fails-in-production-887cb2aaffcf?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/1024/1*ALdjh0cj7aFDmnRQCHTztg.jpeg" width="1024"></a></p><p class="medium-feed-snippet">Building an internal Retrieval-Augmented Generation (RAG) system usually starts as a quick win: you connect a vector database to an LLM&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/why-your-rag-pipeline-works-in-testing-but-fails-in-production-887cb2aaffcf?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/why-your-rag-pipeline-works-in-testing-but-fails-in-production-887cb2aaffcf?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/887cb2aaffcf</guid>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 15:54:18 GMT</pubDate>
            <atom:updated>2026-09-14T15:54:18.634Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Stop Vibe-Checking Your AI Agent: Build a Real Eval Pipeline]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/stop-vibe-checking-your-ai-agent-build-a-real-eval-pipeline-c93bf8514cca?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/2400/1*1wJYVwxpyGot6XnGAJWp-w.png" width="2400"></a></p><p class="medium-feed-snippet">If you have built an agentic system on your own, I have a question for you. How good was the application? Whether your answer is good or&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/stop-vibe-checking-your-ai-agent-build-a-real-eval-pipeline-c93bf8514cca?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/stop-vibe-checking-your-ai-agent-build-a-real-eval-pipeline-c93bf8514cca?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/c93bf8514cca</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[eval-pipeline]]></category>
            <category><![CDATA[ai-evaluation]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Tue, 21 Jul 2026 08:57:30 GMT</pubDate>
            <atom:updated>2026-07-21T08:58:58.050Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Stop Fine-Tuning. You Probably Just Need RAG]]></title>
            <link>https://medium.com/ai-simplified-in-plain-english/stop-fine-tuning-you-probably-just-need-rag-7d0d72d8bd65?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/7d0d72d8bd65</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[agentic-rag]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[llm-finetuning]]></category>
            <category><![CDATA[generative-ai-use-cases]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Thu, 09 Jul 2026 04:49:42 GMT</pubDate>
            <atom:updated>2026-08-01T18:11:02.194Z</atom:updated>
            <content:encoded><![CDATA[<p>Coding agents have shifted where our time goes. We spend less of it wiring up implementations and more of it thinking about the design, how the pieces fit, what the system actually needs to do. That’s the idea, at least. What we see in practice is people handing over the thinking too, not just the typing. The assistant picks the approach, and since it can build almost anything you ask for, it tends to build something big: a fine-tuned model, a vector store, a few moving parts to maintain, for a problem a single well-placed prompt could have handled. RAG versus fine-tuning is exactly the kind of decision that gets skipped this way. It’s worth making on purpose, so this post walks through when each one earns its place.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EZzCMvqbS9MGqYAoBe0z9Q.png" /><figcaption>Image Generated by AI</figcaption></figure><h3>RAG vs Fine-Tuning: How to Choose for a Custom Task</h3><p>A team ships a support chatbot. It works in the demo. Two weeks later, users complain that it quotes an old refund policy that changed last month. One engineer says, “Let’s fine-tune it on our docs.” Another says, “No, we need RAG.” Both are trying to fix the same bug, but they’re proposing very different machines. The choice between RAG and fine-tuning is really a choice about which part of the system you change.</p><h4>What fine-tuning actually does</h4><p>Fine-tuning takes a model that already exists and keeps training it on your own examples. You show it input–output pairs: a question and the answer you want, a ticket and its correct category, a message and the reply in your house style. The training nudges the model’s internal weights so it leans toward those patterns next time.</p><p>Here’s the key point. Fine-tuning changes <strong>how the model behaves</strong>. It’s good at teaching form, tone, and task shape. If you want every answer in one fixed JSON structure, or in a specific brand voice, or sorted into one of eight fixed labels, fine-tuning presses that behavior into the model itself.</p><p>What it doesn’t do well is store fresh facts. The knowledge you train in is frozen at training time. When your refund policy changes, the fine-tuned model won’t know unless you train it again. You also need labeled examples, often many hundreds, plus some compute. Methods like <strong>LoRA</strong> (a parameter-efficient approach that trains a small set of extra weights instead of the whole model) cut the cost a lot, but you still run a training job and host a model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zh7atqniJshf10kko0EExA.png" /><figcaption>Image Generated by AI</figcaption></figure><h4>What a RAG system actually does</h4><p>RAG stands for retrieval-augmented generation. Instead of changing the model, you change what it can see at the moment it answers.</p><p>Here’s the flow.</p><ul><li>You store your documents in a searchable index, usually a vector database that finds text by meaning rather than exact words.</li><li>When a user asks something, the system searches that index, pulls the few most relevant chunks, and pastes them into the prompt as context.</li><li>The model then answers using that material in front of it.</li></ul><p>So RAG changes<strong> what the model knows right now, not how it behaves.</strong> Update a document, and the next answer reflects it, no retraining. You can also show which source each answer came from, which matters when someone needs to trust or audit the result.</p><p>The trade-offs live in retrieval. If the search returns the wrong chunk, the model answers from bad context. Long context costs more per query and can slow the response down. And RAG won’t teach the model a new skill or a new output format, it only feeds it better information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*14G2Vw0L0RsRr6Pl5gfWyQ.png" /><figcaption>Image Generated by AI</figcaption></figure><h3>The one question that decides most cases</h3><p>Before comparing features, ask a plainer question: is your problem about <strong>knowledge</strong> or about <strong>behavior?</strong></p><p>If the model keeps getting facts wrong or out of date, that’s a knowledge problem. RAG is usually the answer.</p><p>If the model knows enough but responds in the wrong shape, the wrong tone, or can’t do the task reliably, that’s a behavior problem. Fine-tuning is usually the answer.</p><h4>When fine-tuning wins</h4><p>Pick fine-tuning when the task is a stable, repeated behavior and you have examples of it done right.</p><p>Some clear cases:</p><ul><li><strong>Strict output format.</strong> You need every response as valid JSON in one schema, every time. A fine-tuned model holds that shape far more reliably than a prompt full of instructions.</li><li><strong>Fixed classification.</strong> Sorting tickets, tagging content, or routing messages into a set list of categories.</li><li><strong>Consistent voice or style.</strong> A tone that’s hard to describe in words but easy to show with a hundred examples.</li><li><strong>Shorter prompts, lower latency.</strong> If you’d otherwise repeat long instructions on every call, baking them in can make each request cheaper and faster.</li><li><strong>A narrow skill the base model does poorly.</strong> A domain’s phrasing, a specialized rewrite task, or a format the model keeps breaking.</li></ul><p>The most common scenario would be when the knowledge isn’t changing much, but the behavior needs to be locked in.</p><h3>When RAG wins</h3><p>Pick RAG when the answer depends on information that lives outside the model, especially if it changes frequently.</p><p>Some clear cases:</p><ul><li><strong>Changing facts.</strong> Product docs, prices, policies, schedules or anything you edit often. RAG reads the current version while a fine-tuned model remembers the old one.</li><li><strong>A large or growing knowledge base.</strong> Thousands of documents you can’t fit into a single prompt or a training set cleanly.</li><li><strong>Answers that need sources.</strong> Legal, medical, finance, internal support cases where you must point to where the answer came from.</li><li><strong>Per-user or per-tenant data.</strong> You can restrict which documents a given user’s query is allowed to retrieve. That’s hard to do once facts are baked into shared weights.</li><li><strong>No labeled training data yet.</strong> You have documents but not clean input–output pairs. RAG lets you start today.</li></ul><p>The most common scenario would be when the behaviour is fine, but the model needs the right, current information in front of it.</p><h3>RAG vs fine-tuning: why the answer is often both</h3><p>These two approaches solve different problems, so real systems frequently use them together.</p><p>Take that support bot from the start. RAG pulls the current help article that answers the user’s question, so the facts are right and up to date. Fine-tuning shapes how the bot replies. The tone, the structure, the way it says “I don’t know” instead of guessing. RAG handles the what while fine-tuning handles the how.</p><p>You don’t have to start with both. Most teams reach for RAG first because it’s faster to stand up and easier to update, then fine-tune later if the response quality or format still isn’t where they need it.</p><p>Well that solves a huge mystery for me. Hope you found the blog insightful.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7d0d72d8bd65" width="1" height="1" alt=""><hr><p><a href="https://medium.com/ai-simplified-in-plain-english/stop-fine-tuning-you-probably-just-need-rag-7d0d72d8bd65">Stop Fine-Tuning. You Probably Just Need RAG</a> was originally published in <a href="https://medium.com/ai-simplified-in-plain-english">AI Simplified 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[A2A Protocol: How AI Agents Delegate Work Without a Human]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/a2a-protocol-how-ai-agents-delegate-work-without-a-human-4940c52b2b7d?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/2600/1*QsfgIKVWUo_9XnMpFDWd4g.png" width="2752"></a></p><p class="medium-feed-snippet">Picture two agents inside the same company. One handles shipping.</p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/a2a-protocol-how-ai-agents-delegate-work-without-a-human-4940c52b2b7d?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/a2a-protocol-how-ai-agents-delegate-work-without-a-human-4940c52b2b7d?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/4940c52b2b7d</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[ai-agents-in-action]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Sun, 21 Jun 2026 08:43:59 GMT</pubDate>
            <atom:updated>2026-06-21T08:43:59.899Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Beyond the Chatbox: Why the Model Context Protocol (MCP) is the “USB Port” for AI]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/beyond-the-chatbox-why-the-model-context-protocol-mcp-is-the-usb-port-for-ai-5dec532436a6?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/1600/1*vtWVkPVXs3R9dNtnLxUQnQ.png" width="1600"></a></p><p class="medium-feed-snippet">In simple terms, MCP is a universal standard that lets AI models safely and easily talk to external tools and data sources.</p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/beyond-the-chatbox-why-the-model-context-protocol-mcp-is-the-usb-port-for-ai-5dec532436a6?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/beyond-the-chatbox-why-the-model-context-protocol-mcp-is-the-usb-port-for-ai-5dec532436a6?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/5dec532436a6</guid>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[model-context-protocol]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[mcps]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Sat, 13 Jun 2026 07:21:55 GMT</pubDate>
            <atom:updated>2026-06-13T07:22:12.693Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Connecting the Dots of AI: From LLMs to Agents and Tools]]></title>
            <link>https://medium.com/@anagharamdas2000/connecting-the-dots-of-ai-from-llms-to-agents-and-tools-5b622d0d57ff?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/5b622d0d57ff</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Wed, 10 Jun 2026 06:12:36 GMT</pubDate>
            <atom:updated>2026-06-10T06:12:36.040Z</atom:updated>
            <content:encoded><![CDATA[<p>The internet is full of tutorials and guides to learn AI. The global market and tech industry saw a major shift in 2022 with the launch of ChatGPT, followed by similar chatbots from competitors. Now, with new models being released every single day, people are struggling to keep up with this ever-changing sector of AI. The uncertainty surrounding this major transition we are experiencing has also induced a sense of panic in people to secure their jobs and future. The fear of being left behind in the AI race has forced professionals to take courses, read tutorials and upskill themselves in this ever-expanding market of AI.</p><p>Professionals in the IT field are familiar with keywords being thrown around like “Agentic AI”, “Image Processing”, “Optical Character Recognition”, “Model Context Protocol”, “Agentic Orchestration” and so on. All these bits and pieces are easy to understand by simply googling or asking your favourite chatbot. We often see people these days type in prompts like “Explain the topic of ‘XYZ’ as if to a 13-year-old…” to understand a topic easily. However, is that enough? With the spike in the consumption of content these days, our capacity to retain information has degraded significantly. You ask Claude or Gemini to explain “Model Context Protocol” as if to a 13-year-old, with examples and analogies, yet a month later you again wonder what the term means. So how does one learn these concepts and retain them in memory?</p><p>The best I could come up with is to understand the connection between all these terms. Learn them by connecting them with a term you already know, and then build on top of it. For example, a normal Computer Science graduate with a beginner’s understanding of AI would know the distinction between Artificial Intelligence, Machine Learning, Deep Learning and Neural Networks. From here, I start connecting the dots for you…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Zvx_AtbHAM41kqhKBhRr6A.png" /><figcaption>Image Generated by AI</figcaption></figure><h3>What are LLMs?</h3><p>Large Language Models, or LLMs, fall into a specific category of AI called Generative AI.</p><p>You might be familiar with the names of certain machine learning models (even though you might not know them in detail) like K Nearest Neighbours, Decision Trees, K-Means Clustering and so on. In a high-level sense, these machine learning models are good at learning patterns from existing data so they can make predictions about new data. Generative AI models, on the other hand, are capable of producing new data based on what they have learned. Large Language Models fall into this category of generative AI models that generate new data based on their training. GPT and Claude are examples of these models.</p><h3>What are AI Agents?</h3><p>Now we move to a new term called Agentic AI. What are AI Agents? Let’s connect the dots from LLMs. Now that you understand LLMs, there is a question I need you to answer. I mentioned that LLMs are models capable of generating new data based on their training. Will they be able to generate data that is different from what they were trained on? After all, they are generating new data.</p><p>The answer is: No. This was a problem during the initial rollout of ChatGPT and other chatbots that followed. They could successfully answer questions about anything within the period they were trained on. For example, if an LLM was trained on data up to 2022, it will not be able to answer the question “Who is the current President of the United States of America?” accurately.</p><p>However, the current versions of these chatbots have solved this problem. The problem was addressed by giving LLMs access to something called “Tools”. These tools give an LLM the capability to fetch data in real time, or from various sources that contain data which was not available at the time of training. For example, an LLM can be given the capability of answering questions based on daily news if it is given a tool that lets it fetch the latest news from the web. Similarly, the same LLM can answer questions about the current weather around the world if it is given another tool that fetches real-time weather data. Do not get stressed by this term “Tool”. Tools are just normal functions you write in code, which are ‘called’ by LLMs.</p><p>An LLM that has one or more tools and can decide on its own which ones to call in order to answer a user’s question is called an Agent. In other words, an AI agent is an LLM with the capability to make its own judgement about which tools to use based on what is needed.</p><p>In this blog, we briefly covered the connection between LLMs, Agents and Tools. Well, that’s three keywords off the checklist. Next time, we’ll connect Tools to the Model Context Protocol, one of the terms we opened with. Keep checking the page for more, and keep connecting the dots!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b622d0d57ff" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Demystifying System Design for Enterprise AI]]></title>
            <link>https://medium.com/@anagharamdas2000/system-design-and-architecture-07ccadcafb2d?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/07ccadcafb2d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[system-design-concepts]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-architecture]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Sat, 28 Feb 2026 10:34:13 GMT</pubDate>
            <atom:updated>2026-02-28T10:35:06.875Z</atom:updated>
            <content:encoded><![CDATA[<p>Imagine you’ve been building cool, quick AI prototypes, a chatbot that answers basic FAQs, or a model that predicts housing prices. Pretty cool, right? Now imagine you need to scale that chatbot to handle thousands of simultaneous customers, or make that housing price model available to an entire bank’s mortgage department.</p><p>That’s where I found myself. As an AI Consultant at a financial firm, my job is to build full-stack applications that use AI to automate complex financial tasks. Think things like analysing regulatory documents, automating data extraction from financial reports, or even helping with risk assessment.</p><p>I was deep into building these applications — stitching together frontend frameworks, setting up backends, and integrating AI models. Then, it hit me. Building the fancy AI part is just the beginning. The real challenge is making it all work together, reliably, and at scale.</p><p>I realised that to be truly effective, I needed to go back to the basics of system design. It was a humbling moment, but a necessary one. I saw that having a strong foundation in system design wasn’t just good to have; it was essential for building enterprise-level applications that not only work once but are robust, scalable, and maintainable.</p><p>In this blog, we’re going to break down system design for enterprise AI, using analogies to make complex concepts simpler.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*apAbhyT2frf1IgCkNPw15g.png" /><figcaption>Image Generated by AI</figcaption></figure><h3>What is System Design?</h3><p>Think of building an enterprise AI application like building a sky-scraper. You can have the most advanced, futuristic elevator system in the world (the AI model). But if the foundation is weak, the plumbing leaks, and the electricity keeps cutting out, no one can actually <em>use</em> that elevator.</p><p><strong>System Design</strong> is like being the architect for that entire sky-scraper. You’re not just designing the elevator; you’re planning:</p><ul><li><strong>Foundation:</strong> Where is everything going to sit? (This is our <strong>Infrastructure &amp; Cloud</strong>).</li><li><strong>Power &amp; Water:</strong> How does everything connect and get the resources it needs? (Our <strong>Backend &amp; APIs</strong>).</li><li><strong>Building Sections:</strong> What goes on which floor? (Our <strong>Architecture Patterns</strong> like Microservices).</li><li><strong>Security Guard:</strong> Who gets to enter, and how is the building protected? (Our <strong>Security</strong> layer).</li><li><strong>Maintenance &amp; Expansion:</strong> How will we fix things if they break, and how can we add more floors later? (Our <strong>Monitoring, Reliability, &amp; Scalability</strong>).</li></ul><p>Let’s break down the essential prerequisites, i.e., the key components required to build a full-fledged system.</p><h4>1. The Strong Foundation: Cloud &amp; Infrastructure</h4><p>Think of the Cloud as renting a supercomputer instead of buying and maintaining one yourself. Companies like AWS (Amazon Web Services), Microsoft Azure, and Google Cloud Platform (GCP) provide huge data centers packed with powerful computers (servers), storage, and networking.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Kh7jogtCK77hmQ4tu5m-vQ.png" /><figcaption>Image Generated by AI</figcaption></figure><ul><li><strong>Scalability:</strong> Just like a restaurant might need more tables during peak hours, your application might need more computing power when more users log in. The Cloud lets you add or remove servers with a few clicks (or even automatically!). This is crucial for AI, which can be computationally expensive.</li><li><strong>Reliability:</strong> Imagine having your own server under your desk. What happens if the power goes out? Or if a pipe bursts? Large cloud providers have multiple data centers across the globe. If one goes down, traffic is automatically routed to another.</li><li><strong>Managed Services:</strong> Instead of setting up and managing a complex database or AI training environment yourself, cloud providers offer pre-built services that handle a lot of the heavy lifting.</li></ul><h4>2. Backend, APIs, &amp; Integrations</h4><p>Your backend is the brain and nervous system of your application. It’s where the actual logic lives — things like processing user requests, talking to the database, and connecting to your AI models.</p><p><strong>Backend </strong>is<strong> </strong>everything that happens on the server after a user clicks a button on the frontend. It’s written in languages like Python (great for AI), Java, Node.js, or Go.</p><p><strong>API (Application Programming Interface)</strong></p><p>Think of an API as a waiter in a restaurant.</p><ul><li><strong>You (The Frontend):</strong> Look at the menu (the user interface).</li><li><strong>The API (The Waiter):</strong> Takes your order (your request) to the kitchen.</li><li><strong>The Kitchen (The Backend / AI Model):</strong> Prepares your food.</li><li><strong>The Waiter (The API):</strong> Brings the finished food back to you.</li></ul><p>APIs allow different parts of your application (like the frontend and the AI model) to talk to each other. Your backend will often expose endpoints (like http://abc.com/xyz) that the frontend can send data to. The backend processes the request and returns a response.</p><h4>3. The Database</h4><p>Think of your database as the memory of your application. It’s a structured way to store all your data, such as user information, application state, and the results from your AI models.</p><p>There are two main types you’ll hear about:</p><ul><li><strong>SQL (Relational Databases):</strong> Picture a massive, perfectly organised Excel spreadsheet. Everything has a fixed structure (rows and columns). These are great for structured data, like user profiles or transaction records. Examples: PostgreSQL, MySQL, SQL Server.</li><li><strong>NoSQL (Non-Relational Databases):</strong> These are like folders filled with different types of documents. They are more flexible and can store various formats without a pre-defined structure. Examples: MongoDB (great for JSON data), DynamoDB.</li></ul><p><strong>For AI, this is huge:</strong> You need to store not just your model but also all the data used to train it and the results it generates. Choosing the right database depends on the type and structure of your data.</p><h4>4. Software Architecture Patterns</h4><p>Now that we have all the pieces, let’s arrange them. This is where architecture patterns come in. Let’s look at two common ones:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xUoL8vsmqEJ86utf86aliA.png" /><figcaption>Image generated by AI</figcaption></figure><ul><li><strong>Monolithic Architecture:</strong> This is like a massive, single-story house. The kitchen, bedrooms, living room, and bathroom are all under one large roof. If you want to expand the kitchen, you might have to structural change the whole house. If the kitchen burns down, the whole house is likely gone. This is how many applications start, but they become hard to manage as they grow.</li><li><strong>Microservices Architecture:</strong> Imagine a small town instead. There’s a bakery (handles one task), a post office (handles another), and a town hall. Each building is independent. If the bakery expands, it only affects the bakery. If one building burns down, the rest of the town keeps running. In enterprise AI, we often use microservices. One service might handle user accounts, another might manage the AI model itself, and another might process the reports. Each can be built, updated, and scaled independently.</li></ul><h4>5. CI/CD and Containers</h4><p>CI/CD (Continuous Integration/Continuous Deployment) comes into play when you need to get your code from your local machine (laptop) to the cloud servers.</p><p>Imagine a highly efficient assembly line in a car factory.</p><ol><li>A robot puts on a door (CI — Continuous Integration: Code from different developers is merged and automatically tested).</li><li>Another robot paints the door (CI: More automated tests).</li><li>The final, tested car gets driven off the line (CD — Continuous Deployment: The tested code is automatically deployed to the production environment).</li></ol><p>This automated pipeline ensures that every change you make is tested and then quickly and reliably made available to users. Tools like GitHub Actions or Jenkins are used for this.</p><p><strong>What about Containers (like Docker)?</strong></p><p>Picture a standard shipping container. Whether it’s loaded onto a ship, a train, or a truck, the contents inside stay exactly the same.</p><p>In software, a container packages up your application plus everything in it to run (libraries, system settings). This ensures that if it works on your laptop, it will work exactly the same way on the production server. This is a game-changer for AI, where complex libraries (like TensorFlow or PyTorch) can often cause head-scratching compatibility issues.</p><h4>6. The Frontend</h4><p>It’s the interface users see and click on — the web pages or mobile app. Modern frontends are built with frameworks like <strong>React</strong> (super popular), Angular, or Vue. These make it easier to build dynamic and responsive user interfaces that can talk to your backend APIs smoothly.</p><h3>Key Considerations When Designing Enterprise Systems</h3><p>Understanding the pillars is great, but there are additional elements to consider when architecting a system.</p><p><strong>1. Scalability (handling growth)</strong></p><p>As your application gets more users, your system needs to grow with them. This is crucial for AI because processing data and running models can be resource-intensive.</p><ul><li><strong>Vertical Scaling:</strong> Buying a faster, more powerful server (like adding a bigger engine to a car). There’s a limit.</li><li><strong>Horizontal Scaling:</strong> Adding more servers (like having multiple cars instead). The cloud excels at this.</li></ul><p><strong>2. Reliability (keeping it running)</strong></p><p>Your system should be fault-tolerant. If a database server fails, is there a backup to take over? If an entire cloud region goes down, can your application automatically switch to another region? This is all about having redundant (backup) components and well-thought-out recovery plans.</p><p><strong>3. Performance (speed matters)</strong></p><p>No one likes a slow application. You need to design for speed. This means things like:</p><ul><li><strong>Caching:</strong> Saving frequently used data (like a pre-calculated AI prediction) in quick-access memory, so you don’t have to re-calculate it every time.</li><li><strong>Asynchronous Processing:</strong> Imagine a long line at a post office. Instead of making you wait, the clerk gives you a ticket and says “We’ll email you when your package is ready.” Your application can do the same. A complex AI task might take time, so the backend can return a “we’re processing it” response immediately and then notify the frontend when it’s done.</li></ul><p><strong>4. Security (protecting everything)</strong></p><p>This is non-negotiable in an enterprise.</p><ul><li><strong>Authentication:</strong> Making sure users are who they say they are (login).</li><li><strong>Authorisation:</strong> Making sure logged-in users only have access to what they’re allowed to (a regular user shouldn’t be able to access financial risk assessment models).</li><li><strong>Data Encryption:</strong> Protecting sensitive data while it’s being transmitted and when it’s being stored in your database.</li></ul><p><strong>5. Monitoring &amp; Observability (knowing what’s happening)</strong></p><p>You can’t fix what you can’t see. Enterprise systems need to be instrumented so you can answer questions like:</p><ul><li>Is the system healthy?</li><li>How many users are online?</li><li>Why is the AI model taking so long to respond?</li><li>Are there any error messages popping up?</li></ul><p>Tools like Prometheus, Grafana, or Datadog are used to collect data (metrics and logs) and visualise it on dashboards, so you can proactively spot and fix problems.</p><h3>Conclusion</h3><p>An enterprise AI application is a complex system, and understanding the core pillars — Cloud, Backend, Databases, Architecture, and Security — is vital for long-term success.</p><p>My journey back to basics was eye-opening. It showed me that the most powerful AI model is useless without a solid system to support it. So, as you start your journey, whether in a university computer science program or through self-teaching, don’t just focus on the fancy new AI frameworks. Build a strong foundation first. Learn how these core components work and play together.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07ccadcafb2d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The 5 Most Critical Security Vulnerabilities in AI Systems: A Beginner’s Guide]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/ai-simplified-in-plain-english/the-5-most-critical-security-vulnerabilities-in-ai-systems-a-beginners-guide-b3bed4fa7660?source=rss-a33659c77ad9------2"><img src="https://cdn-images-1.medium.com/max/2600/1*xz74qJae7FBg5HHccDlIdg.png" width="2816"></a></p><p class="medium-feed-link"><a href="https://medium.com/ai-simplified-in-plain-english/the-5-most-critical-security-vulnerabilities-in-ai-systems-a-beginners-guide-b3bed4fa7660?source=rss-a33659c77ad9------2">Continue reading on AI Simplified in Plain English »</a></p></div>]]></description>
            <link>https://medium.com/ai-simplified-in-plain-english/the-5-most-critical-security-vulnerabilities-in-ai-systems-a-beginners-guide-b3bed4fa7660?source=rss-a33659c77ad9------2</link>
            <guid isPermaLink="false">https://medium.com/p/b3bed4fa7660</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <dc:creator><![CDATA[Anagha Mulloth]]></dc:creator>
            <pubDate>Mon, 02 Feb 2026 06:01:01 GMT</pubDate>
            <atom:updated>2026-02-02T06:02:01.459Z</atom:updated>
        </item>
    </channel>
</rss>