<?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 Hugo Souto, MSc on Medium]]></title>
        <description><![CDATA[Stories by Hugo Souto, MSc on Medium]]></description>
        <link>https://medium.com/@hugosouto?source=rss-eb1012dfd1b5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*7qmlPcJMz-_LtKZhqfpvlQ.png</url>
            <title>Stories by Hugo Souto, MSc on Medium</title>
            <link>https://medium.com/@hugosouto?source=rss-eb1012dfd1b5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 22:11:41 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@hugosouto/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[From Prompt Engineering to Context Engineering]]></title>
            <link>https://medium.com/@hugosouto/from-prompt-engineering-to-context-engineering-8cf1bbbcc153?source=rss-eb1012dfd1b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/8cf1bbbcc153</guid>
            <category><![CDATA[ai-prompt-engineering]]></category>
            <category><![CDATA[ai-context-handling]]></category>
            <category><![CDATA[ai-product-management]]></category>
            <category><![CDATA[retrieval-augmented-gen]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Hugo Souto, MSc]]></dc:creator>
            <pubDate>Sun, 08 Feb 2026 00:45:30 GMT</pubDate>
            <atom:updated>2026-02-08T00:47:34.010Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Expanding the Scope of Enterprise LLM Architecture</strong></h3><p>95% of corporate AI projects fail to deliver measurable returns. That number comes from a <a href="https://sloanreview.mit.edu/article/why-so-many-data-science-projects-fail-to-deliver/">MIT Sloan study</a>, and when we first read it, the reaction is generally: “why so high?”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*71cw7gw1630CzOrOHyCgqA.png" /><figcaption>Prompt Engineering is the tip of the iceberg.</figcaption></figure><p>It’s not for lack of technology. Models keep getting better. It’s not for lack of investment. Companies are pouring billions into AI infrastructure. And it’s not for lack of technical talent. Competent engineers exist in abundance.</p><p>The problem is more subtle. It’s a combination of factors — and data, context, is one of the most neglected, especially with generative AI.<strong>The Myth of the Perfect Prompt</strong></p><p>The industry created a seductive narrative over the past few years: if you master the art of prompt engineering, you’ll have control over AI. Courses emerged, certifications, prompting frameworks. “Prompt Engineer” became a job title at Silicon Valley startups.</p><p>The narrative is attractive because it offers an accessible solution. Write better, get better results. Simple, direct, easy to sell.</p><p>But this narrative ignores a fundamental problem that anyone who’s worked with AI in production knows well: prompts are volatile.</p><p>An LLM’s behavior is a function of the entire context it receives. This includes the prompt, but isn’t limited to it. Moreover, a simple model change often breaks expected behavior:</p><ul><li><strong>Lost context:</strong> what worked on GPT-4 fails on Google’s new model.</li><li><strong>Ignored guardrails:</strong> safety rules stop being respected.</li><li><strong>Instruction drift: </strong>the agent “forgets” instructions that were previously clear.</li></ul><p>Why? Because the prompt is just the tip of the iceberg.</p><h3><strong>Where Projects Actually Die</strong></h3><p>Let me tell you about a case that illustrates the problem well — abstracted to protect confidentiality, but representing a pattern I’ve seen several times in projects at the Brazilian Federal Government’s AI Hub.</p><p>In a large-scale implementation, we had a RAG system connected to a robust knowledge base. The documentation was complete, well-structured, and clearly explained the processes. The context already showed some maturity. It included, for example, a section about special topics users might ask about, like protocol numbers.</p><p>Everything worked well in testing. Users would ask “what’s the status of my protocol number 123XXX7?” and the system found the relevant documentation.</p><p>Until real users started using the system. And most of them didn’t ask complete questions. They simply typed “123XXX7”. Just the number, nothing more.</p><p>The RAG couldn’t connect that input to any document. Why? Because “123XXX7” is just a sequence of numbers. It has no semantic relationship with any content in the knowledge base. There was no vector connecting digits to explanations about protocols.</p><p>The solution wasn’t in the knowledge base. It was in the agent. We needed to teach the agent — via prompt — to recognize protocol patterns before querying the knowledge base. That knowledge needed to live in a different place.</p><p>This case illustrates something already well observed in context engineering theories and other solutions: prompt and context must work together, not compete. But as much as the prompt guides agent behavior, context remains the most important piece, especially in generative AI applications. So we need a discipline for thinking about this in a structured way.</p><h3><strong>What Context Engineering Is</strong></h3><p>Context Engineering is the deliberate design of three fundamental elements:</p><ul><li><strong>What reaches the model:</strong> decisions about which sources to use, how to filter information, what to prioritize and what to exclude. Not everything should feed the model.</li><li><strong>How it’s structured:</strong> information architecture. Long documents or small chunks? Rigid hierarchy or knowledge graph? These choices impact response quality.</li><li><strong>How it’s governed:</strong> maintenance over time. This is where professional implementations differentiate themselves.</li></ul><p><a href="https://www.gartner.com/en/newsroom/press-releases/2024-07-24-gartner-says-60-percent-of-ai-projects-will-be-abandoned-by-2026">Gartner estimates</a> that 60% of AI projects will be abandoned by 2026 due to lack of “AI-ready data.” This isn’t a model problem. It’s a context problem.</p><h3><strong>The Five Markers</strong></h3><p>After several implementations, here are some markers that consistently predict whether a context is ready for production or will die after the demo.</p><blockquote>I’ll cover this topic in a specific article soon.</blockquote><h4><strong>1. Is the context updateable?</strong></h4><p>It seems obvious, but many projects launch with knowledge bases that are never revisited. Without an update cadence and process to detect stale content, the model generates confidence in incorrect answers. Old information is worse than no information. And hallucination isn’t always the model’s fault.</p><h4><strong>2. Does the context have an owner?</strong></h4><p>“It’s everyone’s responsibility” means it’s no one’s. IT manages infrastructure, but who answers for content? Without a clear owner to escalate to when the model fails, the system degrades naturally. And this is where one of the biggest mindset shifts lies: good context engineering often implies changes in organizational information flows, new activities for the newly designated “curators.”</p><h4><strong>3. Is the context versioned?</strong></h4><p>When something breaks — and it always breaks — you need to know what changed. If edits are made without history, there’s no rollback or audit trail. In regulated industries, this becomes a compliance problem. Spreadsheets and traditional documents work to start, systems work to give knowledge bases long life.</p><h4><strong>4. Is the context protected?</strong></h4><p>Does personal data reach the model? Is there security review, encryption, or masking before ingestion? Data leakage through AI is a real risk. Access control must be granular, not “all or nothing.”</p><h4><strong>5. Is the context integrated?</strong></h4><p>This is the most underestimated. If only IT interacts with the knowledge base, it’s born outdated and ages quickly. Living knowledge sits with business teams. If there’s no fluid path for that knowledge to reach the system, context diverges from reality.</p><h3><strong>The Maturity Spectrum</strong></h3><p>These markers form an organizational maturity spectrum:</p><ul><li><strong>Level 1 (Ad-hoc): C</strong>ontext exists without structure. Documents in folders, no owner, sporadic updates.</li><li><strong>Level 2 (Reactive): T</strong>he organization reacts to problems. Fixes happen only when users complain.</li><li><strong>Level 3 (Defined):</strong> Processes exist on paper. Named owner exists, but execution is still irregular.</li><li><strong>Level 4 (Managed):</strong> Processes are followed and measured. There are quality metrics and regular reviews.</li><li><strong>Level 5 (Optimized):</strong> Continuous improvement. Business teams contribute actively and automation detects problems proactively.</li></ul><p>Most organizations invest in cutting-edge models and expensive infrastructure but operate with Level 1 context. It’s like fueling a Formula 1 car with adulterated fuel. You can’t stay in the race.</p><h3><strong>A Quick Check</strong></h3><p>Before continuing, take a quick test with five questions:</p><ol><li>Does your knowledge base have a named owner?</li><li>Are there periodic update mechanisms?</li><li>Can you audit changes and undo wrong edits?</li><li>Is sensitive information filtered before reaching the model?</li><li>Do business teams actively contribute?</li></ol><h4><strong>How to score:</strong></h4><ul><li><strong>5 Yes:</strong> your context is ready to scale.</li><li><strong>3–4 Yes: </strong>solid foundation, but with gaps.</li><li><strong>1–2 Yes:</strong> significant risk — prioritize governance.</li><li><strong>0 Yes:</strong> pause implementation and structure context first.</li></ul><h3><strong>The Perspective Shift</strong></h3><p>Context Engineering doesn’t replace Prompt Engineering. It grounds it.</p><p>We’re in a world increasingly fascinated by ever-larger models. Conferences celebrate new benchmarks. Companies compete for access to the most advanced models. LinkedIn is full of posts about the next prompting framework.</p><p>None of that is wrong. But it’s insufficient.</p><p>In practice, the systems that survive are the most efficient for their purpose. The organizations I’m seeing deliver real value are asking different questions:</p><ul><li>Not just “which model to use?” but also “what reaches the model?”</li><li>Not just “what to put in the prompt?” but also “who maintains the knowledge base?”</li><li>Not just “how to optimize?” but also “how to keep it useful?”</li></ul><p>These questions are less glamorous, but they’re the questions that determine whether your project will deliver value or become part of the 95% that fail.</p><h3><strong>What Comes Next</strong></h3><p>This is the first article in a series on Context Engineering. In the next one, I’ll explore the five dimensions of context architecture — the technical framework that complements the governance markers I presented here.</p><p>📩 And if you need help in your organization, feel free to send me a message, I’m ready to help you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8cf1bbbcc153" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Maximizing Real-World Impact with Machine Learning Models: Two Simple Strategies for Success with…]]></title>
            <link>https://medium.com/@hugosouto/maximizing-real-world-impact-with-machine-learning-models-two-simple-strategies-for-success-with-c5f3d1becea3?source=rss-eb1012dfd1b5------2</link>
            <guid isPermaLink="false">https://medium.com/p/c5f3d1becea3</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automl]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Hugo Souto, MSc]]></dc:creator>
            <pubDate>Wed, 13 Dec 2023 21:16:30 GMT</pubDate>
            <atom:updated>2023-12-13T21:16:30.556Z</atom:updated>
            <content:encoded><![CDATA[<h3>Maximizing Real-World Impact with Machine Learning Models: Two Simple Strategies for Success with AutoML</h3><p>Go beyond notebooks prioritizing solutions to real-world problems</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vlZXXeZJa7dqMCCqIDlFdg.png" /></figure><p>Using notebooks to learn, experiment, develop, and test models is common sense in machine learning. No problem with that unless you want to use your skills in real-world scenarios. In this case, the primary objective of machine learning in business applications is not crafting impressive notebooks but solving tangible, real-world challenges.</p><blockquote>And real-world data are usually bad data.</blockquote><p>In the rapidly evolving field of machine learning, the ability of models to adapt and learn from new data is crucial for applications. This post delves into effective strategies for ensuring your models can be more than experimental exercises but robust solutions to business problems.</p><h3>Adopt AutoML to Fail (And Get It Right) Faster</h3><p>Splitting your data into training, validation, and testing sets is not just a best practice but a necessity to avoid common pitfalls like overfitting. Each set serves a distinct purpose: the training set develops the algorithm, the validation set optimizes hyperparameters and model choice, and the testing set provides an unbiased evaluation of unseen data.</p><blockquote>The best practical advice is to iterate and adapt!</blockquote><p>On the other hand, while speed is critical, ensure the quality of your validation and testing processes. Quality checks at each stage prevent the risks of underfitting or overfitting and secure the model’s applicability in real-world scenarios.</p><p>As machine learning projects can become quite heavy things to move on when they get complex, adopting AutoML in the early stages and cloud-based machine learning tools can be an intelligent choice.</p><p>They work on your training, validation, and test sets, applying different kinds of normalization and other processes in standardized pipelines. It simplifies the process and lets you focus on these numerous but necessary iterations to improve your models and ensure they will work in production.</p><h3>Iterate and Adapt for Model Improvement</h3><p>Regular iterations are essential to refine your models, ensuring they are ready for production and can cope with real-world demands. Adopting AutoML in the early stages can significantly expedite the development process, testing various preprocessing methods and guiding toward effective learning models.</p><p>It helps us to overcome bias in algorithm selection. Comparing many different metrics between various normalization approaches and algorithms lets us be more open to trying new algorithms and technologies. Avoid being biased towards certain methods, and embrace a diverse testing approach to find the most suitable solution.</p><p>In conclusion, prioritize a “fail fast, improve quickly” strategy over perfecting a model that may never see practical use. Concentrate on iterative development to take your project out from scratch, employing AutoML and cloud-based solutions to manage complexity. Second, don’t shy away from experimenting with various algorithms and frameworks, as this exploration is critical to keep the model’s capacity to learn from new data. Like that, you can keep your models not only up-to-date but also highly effective in real-world applications.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c5f3d1becea3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>