<?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 Kannan R on Medium]]></title>
        <description><![CDATA[Stories by Kannan R on Medium]]></description>
        <link>https://medium.com/@kanna.kannanbe?source=rss-11280bddcbae------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Kannan R on Medium</title>
            <link>https://medium.com/@kanna.kannanbe?source=rss-11280bddcbae------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 01:30:48 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kanna.kannanbe/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[AI Engineering | Lesson 6 | Why Choosing the Right AI Model Is More Important Than Building One]]></title>
            <link>https://medium.com/@kanna.kannanbe/ai-engineering-lesson-6-why-choosing-the-right-ai-model-is-more-important-than-building-one-a0e4846e9eab?source=rss-11280bddcbae------2</link>
            <guid isPermaLink="false">https://medium.com/p/a0e4846e9eab</guid>
            <category><![CDATA[software-architecture]]></category>
            <category><![CDATA[java]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Kannan R]]></dc:creator>
            <pubDate>Wed, 08 Jul 2026 15:49:24 GMT</pubDate>
            <atom:updated>2026-07-08T15:49:24.165Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*otGnEkedkjzdpDa7K1vH_Q.png" /></figure><p>When many people first enter the world of AI, they assume that success comes from building a better model. It is easy to understand why. News headlines constantly talk about new models, larger models, and record-breaking benchmarks. As a result, it is natural to think that the biggest challenge in AI is creating the model itself. As I continue learning AI Engineering, I have realized that for most organizations, the challenge is often something entirely different. The real challenge is choosing the right model for the right problem.</p><p><strong>A few years ago, many machine learning projects required teams to train their own models. Today, developers have access to a growing number of foundation models through APIs, open-source releases, and commercial platforms. Instead of asking how to build a model, teams are increasingly asking which model they should use. This shift changes the role of AI Engineering from model creation to model selection and adaptation.</strong></p><p>At first glance, choosing a model may seem straightforward. Why not simply use the model with the highest benchmark score? The problem is that benchmarks do not always reflect real-world requirements. A model that performs exceptionally well on a public leaderboard may not be the best choice for a specific application. Different organizations care about different things. Some prioritize accuracy. Others focus on latency, cost, privacy, or customization. The best model for one company may be completely unsuitable for another.</p><p><strong>Consider a customer support application. One company might value response quality above everything else and be willing to accept higher costs. Another company may need to handle millions of requests every day and therefore prioritize efficiency and cost savings. Both companies are solving similar problems, yet they may choose entirely different models. This illustrates an important lesson: model selection is always tied to business requirements.</strong></p><p>Another factor that influences model selection is data privacy. Many organizations work with sensitive information that cannot easily be shared with external providers. In such cases, using a hosted API may introduce concerns around compliance, governance, or security. Some organizations may decide to host models themselves to maintain greater control over their data. Others may determine that the convenience and capabilities of a managed service outweigh those concerns. There is no universal answer because every organization operates under different constraints.</p><p>Performance is another important consideration. Users expect AI applications to respond quickly. A highly capable model that takes several seconds to answer every request may create a poor user experience. On the other hand, a slightly less capable model that responds almost instantly might be more effective in practice. This reminds me of a common lesson in software engineering: <strong>the technically superior solution is not always the best solution. The best solution is often the one that balances performance, cost, reliability, and maintainability.</strong></p><p>One of the most interesting ideas I encountered is that public benchmarks should be viewed as starting points rather than final answers. Benchmarks can help identify weak models and narrow the list of candidates, but they rarely determine the best choice for a specific application. Real-world evaluation requires testing models against actual business requirements. A model that excels at answering academic questions may perform very differently when generating customer emails, analyzing security incidents, or summarizing internal documents.</p><p>This lesson also reinforced the importance of evaluation. Selecting a model is not a one-time decision. Teams often need to compare multiple models, experiment with prompts, test retrieval strategies, and measure outcomes using meaningful metrics. Success depends less on finding a perfect model and more on creating a systematic process for evaluating options.</p><p>As a software developer, I found this perspective particularly valuable because it shifts the focus from chasing technology trends to solving business problems. The goal is not to use the most advanced model available. The goal is to use the model that best satisfies the requirements of the application. Sometimes that may be the latest frontier model. Other times it may be a smaller, cheaper, and faster alternative.</p><p>Another takeaway is that modern AI Engineering is increasingly about adaptation rather than creation. <strong>Developers spend more time experimenting with prompts, retrieval systems, evaluation pipelines, and application architecture than training foundation models from scratch.</strong> The availability of powerful pre-trained models has dramatically lowered the barrier to entry, allowing teams to focus more on delivering value and less on building core AI capabilities themselves.</p><p>My biggest takeaway from this lesson is simple. The question is no longer “Can we build a model?” The more important question is “Which model best helps us solve the problem?” In many cases, choosing the right model, evaluating it properly, and integrating it effectively will have a far greater impact on the success of an AI application than building a model from scratch. That realization highlights why model selection has become one of the most important skills in modern AI Engineering.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a0e4846e9eab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Engineering | Lesson 5 | Why Building an AI Application Is Easy but Measuring Success Is Hard]]></title>
            <link>https://medium.com/@kanna.kannanbe/lesson-5-why-building-an-ai-application-is-easy-but-measuring-success-is-hard-3b3a29e5b87c?source=rss-11280bddcbae------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b3a29e5b87c</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[aritificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Kannan R]]></dc:creator>
            <pubDate>Fri, 26 Jun 2026 07:58:19 GMT</pubDate>
            <atom:updated>2026-06-26T08:18:01.311Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*otGnEkedkjzdpDa7K1vH_Q.png" /></figure><p>As I continue my journey into AI Engineering, I have started noticing a pattern that surprised me. Most discussions about AI focus on models, prompts, GPUs, or the latest breakthroughs. Very few people talk about something that may actually be more important: evaluation. Before learning about AI Engineering, I assumed that once an AI application was built, its value would be obvious. The reality is much different. Building an AI application is often the easy part. Knowing whether it is actually working is much harder.</p><p>Consider a company that launches an AI-powered customer support chatbot. The chatbot answers customer questions, provides recommendations, and appears to function correctly. Employees are impressed and management is excited. Then someone asks a simple question: Is the chatbot actually helping customers? Suddenly the answer becomes less clear. Are customers getting their problems resolved faster? Are support costs decreasing? Are satisfaction scores improving? Without clear measurements, it becomes difficult to determine whether the application is successful.</p><p>This challenge exists because AI applications are fundamentally different from traditional software. In traditional software development, many outputs are easy to verify. If a program calculates 2 + 2, the expected answer is 4. The result is either correct or incorrect. AI systems operate differently. Imagine asking an AI model to summarize a document. One model may produce Summary A, another may generate Summary B, and a third may create Summary C. All three summaries might be useful even though they are different. There is often no single perfect answer, which makes evaluation much more complicated.</p><p><strong>One of the most valuable ideas I learned is the concept of Evaluation-Driven Development. Instead of building first and measuring later, the process starts by defining how success will be measured. This reminded me of Test-Driven Development in software engineering. In Test-Driven Development, developers write tests before writing code. In AI Engineering, teams should think about evaluation criteria before building the application. By defining success early, teams have a clear way to determine whether their solution is creating value.</strong></p><p>Some applications are naturally easier to evaluate than others. Take fraud detection as an example. Success can often be measured through financial impact. If the system prevents fraudulent transactions and saves money, the value is relatively easy to quantify. Recommendation systems provide another example. Success can be measured through engagement, clicks, purchases, or other business metrics. In both cases, the outcomes can be tracked using clear numbers.</p><p>Customer support chatbots present a different challenge. Multiple factors may influence success, including resolution rate, customer satisfaction, response quality, and escalation rate. None of these metrics alone tells the complete story. Teams often need to evaluate several dimensions at the same time to understand whether the system is performing well.</p><p>I also found it interesting that coding assistants became one of the first highly successful AI applications. One reason is that code is easier to evaluate than many forms of natural language. If an AI generates Java code, the code can be tested. Unit tests either pass or fail. Compilation either succeeds or produces errors. Developers have objective ways to validate the output. This makes coding a relatively friendly environment for AI systems compared to tasks like creative writing or customer conversations.</p><p>This idea led me to another important distinction: closed-ended tasks versus open-ended tasks. Closed-ended tasks usually have a limited number of possible answers. Spam detection is a simple example. An email is either spam or not spam. These tasks are generally easier to evaluate because the expected outcomes are well defined. Open-ended tasks are different. Writing a product description, summarizing a report, or answering customer questions can produce many valid responses. Evaluating quality becomes much more subjective because there is no single correct answer.</p><p>As a Java developer, this lesson changed the way I think about building AI systems. My instinct has often been to focus on implementation details such as architecture, APIs, integrations, and technology choices. While those aspects remain important, AI Engineering introduces another critical question: how will success be measured? A technically impressive system that cannot demonstrate value may ultimately fail regardless of how sophisticated the underlying model is.</p><p>Imagine building a Security Incident Assistant that analyzes firewall logs and provides recommendations. Measuring its effectiveness would require more than checking whether it generates a response. We might evaluate accuracy, completeness, response speed, and actionability. Did it identify the issue correctly? Did it miss important information? Was the recommendation useful enough for an engineer to take action? These questions become just as important as the implementation itself.</p><p><strong>Perhaps the biggest takeaway from this lesson is that many teams spend too much time asking which model they should use and not enough time asking how they will measure success. Models will continue to improve, new capabilities will emerge, and technology will evolve rapidly. However, if an organization cannot define what success looks like, it will struggle to determine whether any AI solution is truly delivering value.</strong></p><p><strong>The more I learn about AI Engineering, the more I realize that successful AI projects are not just about building intelligent systems. They are about building systems that can be evaluated, improved, and trusted. In many cases, the ability to measure success may be more important than the model itself.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b3a29e5b87c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Engineering | Lesson 4 | Why AI Is Becoming Every Developer’s Most Valuable Productivity Tool]]></title>
            <link>https://medium.com/@kanna.kannanbe/lesson-4-why-ai-is-becoming-every-developers-most-valuable-productivity-tool-c6aa94749552?source=rss-11280bddcbae------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6aa94749552</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[java]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <dc:creator><![CDATA[Kannan R]]></dc:creator>
            <pubDate>Thu, 25 Jun 2026 17:09:23 GMT</pubDate>
            <atom:updated>2026-06-26T02:47:07.090Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*otGnEkedkjzdpDa7K1vH_Q.png" /></figure><p>When AI coding assistants first started gaining popularity, many people immediately began asking whether developers would eventually be replaced. As I continue learning AI Engineering, I have started looking at the situation differently. The more I understand how these systems work, the more I believe that the real story is not about replacement. It is about amplification. AI is becoming one of the most powerful productivity tools that developers have ever had access to.</p><p>One reason developers were among the first professionals to adopt AI at scale is that software development is already a language-driven activity. Every day, developers write requirements, documentation, API specifications, SQL queries, code, test cases, design documents, and troubleshooting notes. Large language models are trained on massive amounts of text and code, which makes them naturally suited for many development tasks. Unlike some industries that need to find new ways to use AI, developers can often start benefiting from it immediately.</p><p>What surprised me most is that AI can assist throughout the entire software development lifecycle. It can help generate boilerplate code, explain unfamiliar codebases, create unit tests, suggest refactoring opportunities, write documentation, and even assist with debugging. Tasks that once required significant manual effort can now be completed much faster. <strong>This does not mean the AI understands the business problem better than the developer.</strong> It simply means the developer spends less time on repetitive work and more time on solving meaningful problems.</p><p>A useful way to think about today’s AI tools is to compare them to a junior developer who is incredibly fast but occasionally makes mistakes. The AI can generate code in seconds and provide explanations almost instantly, but it can also produce incorrect solutions, misunderstand requirements, or recommend approaches that are not suitable for a production environment. Because of this, <strong>human judgment remains critical</strong>. The developer is still responsible for reviewing, validating, testing, and deciding whether a solution is appropriate.</p><p>This realization changed how I think about AI adoption. The goal is not to ask whether AI can write code. The more important question is how developers can collaborate with AI to produce better software faster. Just as calculators did not replace accountants and IDEs did not replace programmers, AI is unlikely to eliminate the need for skilled developers. Instead, it changes the way work is performed and increases the importance of higher-level thinking.</p><p>Another interesting observation is that many successful enterprise AI projects are not completely new products. Instead, they improve existing workflows. Organizations are building internal knowledge assistants, documentation generators, code review assistants, incident investigation tools, and support systems that help employees find information more efficiently. In many cases, the value comes from reducing time spent on repetitive tasks rather than creating entirely new capabilities.</p><p>From a business perspective, the impact can be significant. If a developer saves even a small amount of time every day, that improvement scales across entire teams and organizations. A few minutes saved on documentation, testing, debugging, or code generation can add up to hundreds or thousands of hours over the course of a year. This is one reason why companies are investing heavily in AI-powered developer tools.</p><p>One lesson I continue to see throughout AI Engineering is that technology alone rarely creates value. The value comes from how people use it. A powerful language model is impressive, but its true impact appears when it becomes part of a workflow that helps individuals and teams accomplish their goals more effectively. Developers who learn how to work alongside AI will likely gain a significant advantage, not because the AI replaces their expertise, but because it allows them to apply that expertise more efficiently.</p><p>My biggest takeaway from this lesson is simple. <strong>The future of software development is not developers versus AI. It is developers with AI.</strong> Those who learn how to use these tools effectively will spend less time on routine tasks and more time focusing on architecture, business problems, system design, and innovation. In that sense, AI may become one of the most valuable productivity tools a developer can have.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6aa94749552" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Engineering | Lesson 3 | The Three AI Engineering Skills Every Developer Should Learn First]]></title>
            <link>https://medium.com/@kanna.kannanbe/lesson-3-the-three-ai-engineering-skills-every-developer-should-learn-first-40e9d64b8f43?source=rss-11280bddcbae------2</link>
            <guid isPermaLink="false">https://medium.com/p/40e9d64b8f43</guid>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[java]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Kannan R]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 08:43:42 GMT</pubDate>
            <atom:updated>2026-06-26T02:47:23.994Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*otGnEkedkjzdpDa7K1vH_Q.png" /></figure><p>As I continue learning AI Engineering, I keep discovering something surprising: building AI applications is often less about training models and more about using existing models effectively. Before starting this journey, I assumed AI engineers spent most of their time building neural networks and training massive models. The more I learn, the more I realize that most real-world AI projects focus on something else entirely.</p><p><strong>One of the biggest lessons so far is that there are three core techniques behind many modern AI applications: Prompt Engineering, Retrieval-Augmented Generation (RAG), and Fine-Tuning. Understanding the difference between these three concepts helped me see how companies actually build AI products.</strong></p><p>The first technique is Prompt Engineering. This is the easiest and most accessible way to improve AI output. Instead of changing the model itself, you change the instructions you give it. Think about asking a colleague for help. If you simply say, “Analyze this log,” you may get a vague answer. If you say, “You are a cybersecurity analyst. Classify this log as a security incident, configuration issue, or network failure and explain your reasoning,” the response will usually be much better. The same principle applies to large language models. Better instructions often produce better results.</p><p>As a Java developer, I found this idea interesting because it feels similar to designing APIs. A poorly designed API request usually produces poor results. A well-designed request produces predictable and useful results. Prompt Engineering is essentially learning how to communicate effectively with AI systems.</p><p>The second technique is Retrieval-Augmented Generation, commonly known as RAG. This concept answered a question I had for a long time: how can an AI system answer questions about information that didn’t exist when the model was trained?</p><p>Imagine a company has internal security policies, architecture documents, and operational procedures. A public language model has never seen those documents. One option would be retraining the model every time a document changes, but that would be expensive and impractical. RAG solves this problem by retrieving relevant information first and then providing that information to the model along with the user’s question.</p><p>This approach feels very familiar to anyone with backend development experience. In a traditional application, we query a database before returning a response. In a RAG application, we retrieve relevant knowledge before asking the model to generate an answer. The architecture is different, but the underlying idea is surprisingly similar.</p><p>What makes RAG powerful is that company knowledge can change without requiring model retraining. Update the documents and the AI assistant immediately has access to the latest information. This is one reason why so many enterprise AI solutions are built using RAG.</p><p>The third technique is Fine-Tuning. While prompts and RAG provide additional instructions and knowledge, Fine-Tuning actually changes the behavior of the model itself. This is useful when an organization wants highly specialized responses or a consistent communication style.</p><p>For example, imagine a company has thousands of high-quality incident reports written by experienced security analysts. The organization may want every AI-generated report to follow the same structure, tone, and terminology. In cases like this, Fine-Tuning can help the model learn those patterns.</p><p>One of the most valuable lessons from this chapter is that Fine-Tuning should not be the first solution. Many developers immediately assume they need to train or retrain models. In reality, the recommended order is usually much simpler. Start with Prompt Engineering. If the model lacks knowledge, add RAG. Only consider Fine-Tuning when the first two approaches are no longer sufficient.</p><p>This idea reminds me of software engineering best practices. We don’t immediately build a distributed system for every problem. We start with the simplest solution and increase complexity only when necessary. AI Engineering follows a similar mindset.</p><p>Another important takeaway is that the availability of foundation models has changed the role of engineers. In the past, many machine learning projects focused on creating models. Today, many AI projects focus on adapting existing models to solve business problems. The challenge is no longer building intelligence from scratch. The challenge is integrating intelligence into products and workflows.</p><p>As someone coming from a Java Developer background, I find this encouraging. Skills such as API design, system architecture, microservices, security, and database integration are still highly relevant. The new skills I need to learn are Prompt Engineering, RAG, vector databases, evaluation techniques, and AI application design.</p><p>My biggest takeaway from this lesson is simple: successful AI applications are rarely powered by the model alone. The real value comes from how we combine prompts, knowledge retrieval, and application logic to solve real business problems. The model is important, but the engineering around the model is what transforms a powerful technology into a useful product.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=40e9d64b8f43" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Engineering | Lesson 2 | The Breakthrough Behind ChatGPT: Why Large Language Models Suddenly…]]></title>
            <link>https://ai.plainenglish.io/lessson-2-the-breakthrough-behind-chatgpt-why-large-language-models-suddenly-became-possible-143e1a5459ec?source=rss-11280bddcbae------2</link>
            <guid isPermaLink="false">https://medium.com/p/143e1a5459ec</guid>
            <category><![CDATA[java]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Kannan R]]></dc:creator>
            <pubDate>Wed, 24 Jun 2026 04:35:17 GMT</pubDate>
            <atom:updated>2026-06-26T02:46:36.102Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI Engineering | Lesson 2 | The Breakthrough Behind ChatGPT: Why Large Language Models Suddenly Became Possible</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*otGnEkedkjzdpDa7K1vH_Q.png" /></figure><p>As I continue my journey into AI Engineering, one question kept bothering me:</p><p>If language models are fundamentally just predicting the next word, why didn’t ChatGPT exist 10 or 20 years ago?</p><p>The second lesson from my AI Engineering learning journey helped answer that question. The answer isn’t a single breakthrough. It is the combination of three things: data, scale, and a new way of learning.</p><h3>The Problem with Traditional Machine Learning</h3><p>Before modern foundation models, most machine learning systems depended heavily on labeled data.</p><p>For example, if we wanted to build a sentiment analysis system, we needed examples like:</p><ul><li>“Great product” → Positive</li><li>“Poor quality” → Negative</li></ul><p>Someone had to create those labels.</p><p>This works, but it creates a bottleneck. Collecting millions of labels is expensive, time-consuming, and often impractical.</p><p>As a software engineer, I can relate this to manually maintaining configuration data. It works at small scale, but eventually the maintenance effort becomes the real problem.</p><h3>The Genius Idea: Let the Data Label Itself</h3><p>Modern language models use something called self-supervised learning.</p><p>Instead of relying on humans to provide labels, the training process creates labels automatically.</p><p>Consider a sentence:</p><p>“I love playing football on weekends.”</p><p>A model can hide one word and try to predict it.</p><p>The sentence itself becomes both the question and the answer.</p><p>Suddenly, every document on the internet becomes training data.</p><p>No human annotation required.</p><p>This idea seems simple, but it completely changes the economics of AI.</p><p>Instead of being limited by labeled datasets, models can learn from books, websites, research papers, code repositories, and countless other sources.</p><p>The internet effectively becomes the training dataset.</p><h3>Why Are They Called Large Language Models?</h3><p>The word “large” is not marketing language.</p><p>It refers to scale.</p><p>Modern language models became possible because three things grew dramatically.</p><h3>1. More Data</h3><p>Earlier machine learning projects often trained on thousands or millions of examples.</p><p>Today’s language models learn from trillions of tokens.</p><p>The amount of available text on the internet provides a learning resource that previous generations of AI simply didn’t have.</p><h3>2. More Parameters</h3><p>Parameters are the internal values that a model learns during training.</p><p>A useful mental model is to think of parameters as the model’s capacity to capture patterns.</p><p>Earlier models contained millions of parameters.</p><p>Modern frontier models contain billions or even trillions.</p><p>This increase in scale allows models to learn far more complex relationships.</p><h3>3. More Computing Power</h3><p>Training modern AI systems requires enormous computational resources.</p><p>Thousands of GPUs may be used during training.</p><p>The combination of abundant data, larger models, and powerful hardware created the conditions necessary for the current AI revolution.</p><h3>The Shift from Single-Purpose Models to Foundation Models</h3><p>One of the most interesting ideas I learned is the difference between traditional machine learning models and foundation models.</p><p>Historically, many AI systems were built for a single task.</p><p>One model might detect spam.</p><p>Another model might translate languages.</p><p>A third model might classify documents.</p><p>Each problem often required a different solution.</p><p>Foundation models changed that approach.</p><p>A single model can now perform many different tasks:</p><ul><li>Translation</li><li>Summarization</li><li>Question answering</li><li>Classification</li><li>Content generation</li><li>Coding assistance</li></ul><p>Instead of building a new model for every business problem, developers can adapt one powerful model to many different use cases.</p><p>This is one of the reasons AI Engineering has emerged as a discipline of its own.</p><h3>Foundation Models and LLMs Are Not Exactly the Same</h3><p>Before this lesson, I used the terms interchangeably.</p><p>Now I understand that they are related but different.</p><p>A Large Language Model (LLM) primarily works with text.</p><p>A Foundation Model is a broader concept.</p><p>Foundation models can work with:</p><ul><li>Text</li><li>Images</li><li>Audio</li><li>Video</li><li>Code</li></ul><p>In other words:</p><p>Every LLM is a foundation model.</p><p>Not every foundation model is an LLM.</p><p>This distinction becomes increasingly important as AI systems become more capable.</p><h3>The Rise of Multimodal AI</h3><p>Humans do not understand the world through text alone.</p><p>We see images.</p><p>We hear sounds.</p><p>We watch videos.</p><p>Modern AI systems are evolving in the same direction.</p><p>Instead of accepting only text, newer models can process multiple forms of information simultaneously.</p><p>For example:</p><ul><li>Upload a network architecture diagram and ask for an explanation.</li><li>Upload an application error screenshot and request troubleshooting guidance.</li><li>Upload a PDF and ask questions about its contents.</li></ul><p>This ability to combine different information types is known as multimodal AI.</p><p>As someone coming from a backend engineering background, I find this particularly exciting because future enterprise applications will likely combine documents, screenshots, logs, diagrams, and natural language conversations within a single workflow.</p><h3>My Biggest Takeaway</h3><p>The most important insight from this lesson is that the AI revolution did not happen because someone invented a completely new idea.</p><p>The underlying concept of predicting the next token is surprisingly simple.</p><p>What changed was scale.</p><p>More data.</p><p>More parameters.</p><p>More compute.</p><p>Combined with self-supervised learning, these factors unlocked capabilities that were previously impossible.</p><p>As software engineers, we are now in a unique position.</p><p>We don’t need to train billion-parameter models ourselves.</p><p>We can focus on building products and business solutions on top of foundation models.</p><p>That shift is what makes AI Engineering such an exciting field today.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=143e1a5459ec" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/lessson-2-the-breakthrough-behind-chatgpt-why-large-language-models-suddenly-became-possible-143e1a5459ec">AI Engineering | Lesson 2 | The Breakthrough Behind ChatGPT: Why Large Language Models Suddenly…</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[AI Engineering | Lesson 1 | My First Lesson in AI Engineering: Why You Don’t Need to Build GPT to…]]></title>
            <link>https://ai.plainenglish.io/my-first-lesson-in-ai-engineering-why-you-dont-need-to-build-gpt-to-build-ai-products-439e3a9dc5c8?source=rss-11280bddcbae------2</link>
            <guid isPermaLink="false">https://medium.com/p/439e3a9dc5c8</guid>
            <category><![CDATA[ai-engineering]]></category>
            <category><![CDATA[java]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Kannan R]]></dc:creator>
            <pubDate>Fri, 19 Jun 2026 05:36:33 GMT</pubDate>
            <atom:updated>2026-06-26T02:45:56.663Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI Engineering | Lesson 1 | My First Lesson in AI Engineering: Why You Don’t Need to Build GPT to Build AI Products</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*otGnEkedkjzdpDa7K1vH_Q.png" /></figure><h3>My First Lesson in AI Engineering: Why You Don’t Need to Build GPT to Build AI Products</h3><p>Over the past few years, Artificial Intelligence has evolved from a niche research topic into something that almost every software engineer can use in their daily work. As a Java developer looking to transition into AI Engineering, I recently started reading <em>AI Engineering: Building Applications with Foundation Models</em> by Chip Huyen. The very first lesson challenged many of my assumptions about AI and completely changed the way I think about building intelligent systems.</p><p>Before learning about AI Engineering, I assumed that building AI products meant creating machine learning models from scratch. Traditionally, machine learning projects followed a fairly complex process. Teams had to collect data, clean and label it, train a model, deploy it into production, and then continuously maintain and improve it. This approach certainly works, but it requires significant expertise, infrastructure, and time. Building a production-ready machine learning model could take months of effort. For many organizations, the biggest challenge wasn’t finding ways to use AI. The real challenge was building AI itself.</p><p>Today, things are very different. We now have powerful foundation models such as GPT, Claude, Gemini, and many other large language models. These models have already been trained on enormous amounts of data and can perform a wide range of tasks. Instead of spending months training a model from scratch, developers can simply call an API and build applications on top of these existing models.</p><p>This shift reminded me of the evolution of cloud computing. Years ago, companies had to purchase physical servers, manage hardware, and maintain infrastructure themselves. Then cloud providers changed everything by offering infrastructure as a service. Foundation models are doing something very similar for AI. Instead of building the intelligence yourself, you can access intelligence as a service and focus on solving business problems. This simple idea is at the heart of AI Engineering.</p><p>So what exactly is AI Engineering? After reading the first lesson, I would describe it as the practice of building applications using existing foundation models. Rather than spending most of the effort on training models, AI Engineers focus on areas such as Prompt Engineering, Retrieval-Augmented Generation (RAG), Fine-Tuning, Application Integration, Evaluation, and Monitoring. In other words, the emphasis shifts from model creation to product creation.</p><p>As a Java developer, I found this particularly exciting because many of the skills I have already developed remain highly valuable. Building APIs, designing scalable systems, handling security, managing databases, and creating reliable services are still critical skills. AI simply becomes another component within the overall architecture. Instead of replacing traditional software engineering, it extends it.</p><p>One of the most interesting concepts in the lesson was understanding how language models actually work. When we interact with systems like ChatGPT, it often feels like they are thinking or reasoning in a human-like way. However, at their core, language models are prediction machines. Their primary job is surprisingly simple: predict the next token.</p><p>For example, consider the sentence:</p><p>“My favorite color is ___”</p><p>The model predicts the most likely continuation based on patterns it learned during training. It does not think the way humans think. It does not reason exactly as we do. Instead, it learns statistical relationships between pieces of text and uses those relationships to generate responses. Although the idea sounds simple, the results can be remarkably powerful.</p><p>Another concept that caught my attention was the idea of tokens. Before reading about language models, I assumed that models processed complete words. That assumption turned out to be incorrect. Most language models work with tokens rather than words.</p><p>A token can be:</p><ul><li>A word</li><li>Part of a word</li><li>A punctuation mark</li><li>A special symbol</li></ul><p>For example, the word:</p><p>“running”</p><p>might be split into:</p><ul><li>run</li><li>ning</li></ul><p>This approach helps models reduce vocabulary size, handle new words more effectively, and learn reusable language patterns. Understanding tokens is important because model costs, context windows, and performance limits are often measured in tokens rather than words.</p><p>The lesson also introduced two important types of language models: BERT and GPT. Understanding the difference between them helped me better appreciate how modern AI systems work.</p><p>BERT can be thought of as a “fill in the blank” model. It is trained to predict missing words within a sentence. For example:</p><p>“My favorite ____ is blue.”</p><p>BERT looks at both the left and right side of the missing word to determine what belongs in the blank. Because of this capability, BERT became widely used for tasks such as search, classification, and sentiment analysis.</p><p>GPT works differently. Instead of filling in missing words, GPT focuses on continuing text. Given a prompt such as:</p><p>“My favorite color is”</p><p>the model predicts the next token. Then it predicts the next token after that. Then another. This process continues until a complete response is generated.</p><p>The easiest way I found to remember the difference is:</p><p><strong>BERT understands. GPT generates.</strong></p><p>Perhaps the biggest insight from this lesson was understanding why Generative AI has become so powerful. Many tasks that seem completely different can actually be reframed as text generation problems.</p><p>Translation becomes text generation.</p><p>Summarization becomes text generation.</p><p>Classification becomes text generation.</p><p>Question answering becomes text generation.</p><p>Coding becomes text generation.</p><p>Once a model becomes extremely good at predicting the next token, it suddenly becomes useful for a vast number of tasks. This is one of the key reasons foundation models have transformed the software industry so rapidly.</p><p>My biggest takeaway from this lesson is that becoming an AI Engineer does not mean becoming an AI researcher. Most companies do not need another GPT. They need engineers who can use existing foundation models effectively to solve real business problems.</p><p>For software engineers, this is encouraging news. The focus shifts from training models to designing intelligent applications. Instead of spending years learning how to build large language models from scratch, we can focus on understanding how to integrate them into products, create better user experiences, and deliver business value.</p><p>That feels like a much more practical and accessible path into AI. And after completing just the first lesson, I am even more excited to continue this AI Engineering journey and learn what comes next.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=439e3a9dc5c8" width="1" height="1" alt=""><hr><p><a href="https://ai.plainenglish.io/my-first-lesson-in-ai-engineering-why-you-dont-need-to-build-gpt-to-build-ai-products-439e3a9dc5c8">AI Engineering | Lesson 1 | My First Lesson in AI Engineering: Why You Don’t Need to Build GPT to…</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>
    </channel>
</rss>