<?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 Alex Worapol Pongpech on Medium]]></title>
        <description><![CDATA[Stories by Alex Worapol Pongpech on Medium]]></description>
        <link>https://medium.com/@alexpongpech?source=rss-6ec1802f3d5c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*aVyo2SC2N1ELoVezc2cghA.jpeg</url>
            <title>Stories by Alex Worapol Pongpech on Medium</title>
            <link>https://medium.com/@alexpongpech?source=rss-6ec1802f3d5c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:04:55 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@alexpongpech/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[Your Data Platform Is Not a Semantic Layer]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-snippet">Agents will expose the difference faster than any dashboard ever did.</p><p class="medium-feed-link"><a href="https://medium.com/@alexpongpech/your-data-platform-is-not-a-semantic-layer-0a6d0802baa8?source=rss-6ec1802f3d5c------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@alexpongpech/your-data-platform-is-not-a-semantic-layer-0a6d0802baa8?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/0a6d0802baa8</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 12:43:21 GMT</pubDate>
            <atom:updated>2026-11-05T12:43:21.498Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Time Is the One Thing I Can’t Scale]]></title>
            <link>https://medium.com/@alexpongpech/time-is-the-one-thing-i-cant-scale-136c422edc79?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/136c422edc79</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 01:26:29 GMT</pubDate>
            <atom:updated>2026-10-15T01:26:29.782Z</atom:updated>
            <content:encoded><![CDATA[<p>I knew going into the role of Chief Technology and AI Officer that the technology would be hard. Platforms, AI strategy, architecture decisions that stick around long after you’ve moved on. None of that was a surprise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZlMmFBJL8kMhN77l6Dw3KQ@2x.jpeg" /><figcaption>Photo by Nathan Dumlao on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I also knew that time would be the real constraint. I’ve believed that for most of my career. What the role has done is reconfirm it, every single week, more clearly than any job before it.</p><h4>How a week disappears</h4><p>Most days, something wants a piece of me. A meeting that probably could have been a quick message. A request that feels urgent to the person sending it but doesn’t have much to do with what we’re actually trying to build. Sometimes it’s a good idea, a genuinely good one, that just isn’t the right thing to work on now.</p><p>None of it is bad. That’s the trap. Every request makes sense on its own. It’s only when you look back on Friday afternoon that you see how easily a week can be busy every hour and still move nothing that matters.</p><p><em>I work in a field obsessed with scale. We scale infrastructure. We scale teams. We talk about scaling AI across the whole enterprise. But nobody has found a way to scale a day. Still twenty-four hours. No diagram changes that.</em></p><h4>What “no” actually means</h4><p>Saying no is not a rejection. When I turn down something that doesn’t need me, I’m protecting time for things that do. My team, for one. They need me paying attention to the decisions that are blocking them, not half-listening while I answer emails. There are also problems that only someone in my seat can solve. And there’s the slow thinking that real strategy needs, the kind that never shows up in a fifteen-minute gap between calls.</p><p>So a no isn’t really about the person asking. It’s about the priorities we’ve already agreed on. When I explain it like that, people usually get it. Some of them look relieved, to be honest, because they’re buried under the same pile of small asks.</p><h4>You have to know your yes first</h4><p>Saying no doesn’t help much if you haven’t decided what’s important. You free up an hour, and the next urgent-but-not-important thing walks in and takes it. Time you protect without a purpose just sits there waiting to be claimed.</p><p><strong>The real work is deciding ahead of time what deserves your hours</strong>. Once that’s clear, no gets easier. You stop agonizing over every request. You just ask one question: does this serve what we’ve already said matters? Most of the time the answer is obvious.</p><h4>You can manage time. You can’t control it.</h4><p>There’s a popular idea that if you plan carefully enough, your week will go the way you designed it.</p><p>It doesn’t. It never has.</p><p>Time moves. Something breaks in production on a Tuesday morning. A client escalation reshuffles everything. A decision that seemed closed comes back open. Something that was the top priority last month slowly becomes less important this month, often without anyone saying so.</p><p>So I don’t try to control time. I manage it, and those are really different things. Control assumes everything holds still. Managing something means accepting that it’s alive and keeps shifting, and that it needs your attention all the time.</p><p>In practice, I treat my time a bit like we treat any critical system at work. I keep an eye on it. I look at where my hours actually went, not where I planned for them to go. I try to catch the drift early, when one type of request starts creeping in and crowding out the work that matters. And I adjust. Sometimes daily. Not once a quarter when it’s already too late.</p><p>I’m not chasing a perfect calendar. I want one that responds. The leaders I respect most aren’t the ones whose plans never change. They’re the ones who notice the change early and rebalance before the important things get pushed aside.</p><h4>It’s the same outside of work</h4><p>This applies to life, not just the job.</p><p>Health doesn’t happen in whatever time is left over. If you don’t protect it, there’s nothing left. Same with family. Same with rest, and with those unplanned hours where clarity seems to show up on its own. These things are always important and almost never urgent. That’s exactly why they lose to everything else unless you guard them on purpose.</p><p>Life lasts as long as the time we have. Our health reflects how much of that time we give it. People hear lines like that as motivational posters. I’ve always seen them as arithmetic.</p><h4>What AI can’t do for us</h4><p>A lot of my work is about artificial intelligence and what it can do for organizations. And it can do a lot. It makes us faster. It takes repetitive work off our plates. It helps us see things we’d miss on our own.</p><p>But it can’t give us more time. At best it hands some time back, and what we do with that is still completely up to us. If we fill every saved hour with more low-value work, we’ve just automated ourselves into the same busy, empty week, only faster.</p><p><strong>That’s the real leadership question right now. Not how do we do more, but how do we choose better.</strong> Technology amplifies our choices. It doesn’t make them for us.</p><h4>Where I stand</h4><p>None of this is new to me. But it’s worth saying out loud, because it’s easy to know something and still let the calendar win. I don’t judge a good week by how full it was. I judge it by whether the hours went to what matters.</p><p>Every Monday I ask myself two questions. What’s one thing I’ll say no to this week? And where has my time been drifting without my noticing?</p><p>I’d be curious how you’d answer them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=136c422edc79" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why We’re Betting on OKRs to Align a 300-Person Tech and AI Organization]]></title>
            <link>https://medium.com/@alexpongpech/why-were-betting-on-okrs-to-align-a-300-person-tech-and-ai-organization-00d51a7de89c?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/00d51a7de89c</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 00:49:49 GMT</pubDate>
            <atom:updated>2026-09-22T00:49:49.826Z</atom:updated>
            <content:encoded><![CDATA[<p>When a technology organization is small, alignment happens by accident. Everyone sits near each other, hears the same conversations, and knows what matters this quarter. At around 300 people, spread across platform engineering, infrastructure, applications, data, and AI, that accident stops happening. Teams get busy, roadmaps multiply, and “we’re all working hard” quietly replaces “we’re all working on the right things.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cr8KcPmkXkL-topt" /><figcaption>Photo by <a href="https://unsplash.com/@startaeteam?utm_source=medium&amp;utm_medium=referral">Startaê Team</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>That’s the problem we’re solving with Objectives and Key Results (OKRs). This article explains how we’re adapting them for an organization that spans both a traditional technology domain and an AI domain, and the mistakes we’re deliberately trying to avoid.</p><h3>What OKRs actually are</h3><p>The structure is simple. An OKR pairs an objective, a significant and clearly defined goal, with three to five key results that serve as measurable success criteria for that goal. Objectives can also be supported by initiatives, which are the plans and activities that move the key results forward.</p><p>The distinction between those three layers is where most teams go wrong, so here is how we frame it internally. The objective is <em>where we want to go</em>, and it should be qualitative and worth caring about. The key results are <em>how we’ll know we got there</em>, and they must be measurable outcomes, not tasks. The initiatives are <em>what we’ll do</em>, and they belong in your backlog, not in your OKRs.</p><p><strong>A key result that reads “Launch the governance engine” is an initiative wearing a disguise. A real key result reads “Ensure 100% of production AI agents run under automated policy enforcement, while cutting governance review time from three weeks to three days.”</strong> The first can be completed while delivering zero value. The second can’t. The first can be completed while delivering zero value. The second can’t.</p><h3>Why OKRs, and not just more KPIs?</h3><p>We already track plenty of KPIs: uptime, incident counts, delivery velocity, cloud spend. KPIs tell you whether the machine is healthy. OKRs tell you where the machine should be heading next.</p><p>We think of it this way: KPIs are the dashboard, and OKRs are the destination. A team can have perfectly green KPIs while making no progress on the strategic shifts the business actually needs. <em>OKRs force the conversation about change, not just maintenance.</em></p><p>There’s also a cultural difference. Older management-by-objectives systems encouraged people to set safe goals because hitting 100% was what got rewarded; OKRs deliberately broke that link so teams could aim higher without fear of being punished for missing. The commonly recommended target success rate for key results is around 70%. If a team hits 100% on everything, their goals were too comfortable.</p><h3>How we’re adapting OKRs for a tech and AI organization</h3><h3>Two domains, one direction</h3><p>Our organization spans a technology domain (platforms, infrastructure, applications, reliability) and an AI domain (models, agentic systems, AI engineering, governance). These groups work on different time horizons and measure success differently. Infrastructure work often shows value in stability and cost curves; AI work often shows value in experimentation speed and adoption.</p><p><strong>We don’t force them into identical OKRs. Instead, both domains roll up to a small number of shared organizational objectives, and each domain writes key results in its own language. Alignment means pointing in the same direction, not using the same metrics.</strong></p><h3>Fewer objectives, on purpose</h3><p><strong><em>At the organization level, we limit ourselves to three to four objectives per cycle</em></strong>. <strong><em>At the team level, two to three.</em></strong> Every additional objective dilutes attention. If everything is a priority, nothing is.</p><h3>Quarterly cycles, annual intent</h3><p>Our organizational objectives are set with a yearly view, but key results are written and scored quarterly. Technology, and AI especially, moves too fast for annual key results to stay meaningful.</p><h3>Examples of what good looks like</h3><p><strong>For the technology domain, an objective might be <em>“Make our platform the fastest, safest path to production.</em></strong><em>”</em> Supporting key results could include reducing median deployment lead time to under one day, cutting change failure rate below a set threshold, and getting a target share of teams onto the standard delivery pipeline.</p><p><strong>For the AI domain, an objective might be <em>“Move AI from pilots into production value.”</em></strong> Supporting key results could include increasing the number of AI use cases running in production with measured business outcomes, reducing time from approved use case to first production release, and ensuring every production AI system passes governance and risk review before launch.</p><p>Notice what’s absent: “Hire five ML engineers,” “Build a model registry,” “Evaluate three vendors.” Those may be necessary, but they’re initiatives. They go in the plan, not the scorecard.</p><h3>The mistakes we’re deliberately avoiding</h3><p><strong>Tying OKRs directly to performance ratings.</strong> The moment a missed key result costs someone their bonus, people stop setting ambitious goals. We keep OKR scoring separate from individual performance evaluation. Performance conversations can reference OKRs as context, but a 0.6 score on a stretch goal is not a failure.</p><p><strong>Cascading everything top-down.</strong> With 300 people, a purely top-down cascade becomes a telephone game and kills ownership. Leadership sets the organizational objectives; teams propose their own key results that ladder up. The best key results usually come from the people closest to the work.</p><p><strong>Writing OKRs for every individual.</strong> At our size, individual OKRs create administrative overhead with little benefit. <strong>We set OKRs at the organization, domain, and team levels, and stop there.</strong></p><p><strong>Set-and-forget.</strong> An OKR reviewed only at quarter end is just a report card. Teams check in on confidence levels every two weeks, and leadership reviews progress monthly. The point of a check-in isn’t to report status; it’s to surface blockers early enough to do something about them.</p><p><strong>Treating the framework as the goal.</strong> Nobody gets promoted for writing beautiful OKRs. If the process becomes heavier than the value it creates, we simplify it.</p><h3>What we expect to get wrong</h3><p>We’ll almost certainly write some bad key results in our first cycle. Some will be tasks in disguise, some will be unmeasurable, and some will be too safe. That’s expected. Most organizations need two to three cycles before OKRs start working as intended, and the retrospective at the end of each quarter matters as much as the planning at the start.</p><p>What we’re really building isn’t a goal-setting system. It’s a habit: regularly asking, as an entire organization, <em>what matters most right now, and how will we know if we’re making progress?</em></p><p>At 300 people, that question no longer answers itself. OKRs are how we make sure it gets asked.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=00d51a7de89c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Don’t Let Your AI Agents Touch Your Data Before They Touch the Semantic Model]]></title>
            <link>https://medium.com/@alexpongpech/dont-let-your-ai-agents-touch-your-data-before-they-touch-the-semantic-model-5071a0983ac3?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/5071a0983ac3</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Fri, 18 Sep 2026 00:46:01 GMT</pubDate>
            <atom:updated>2026-09-18T00:46:01.585Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The shortest path from question to answer is usually the most expensive one</em></p><p>My meeting was just cancelled, so I pen this. This issue has been bothering me quite a bit lately. Picture a familiar demo. Someone types “What was our revenue last quarter by region?” into a chat window. A few seconds later, an AI agent writes a SQL query, runs it against the warehouse, and returns a clean table with a chart. People nod.</p><p>Then someone from finance asks one question: “Which revenue?”</p><p>Gross bookings? Recognized revenue? Net of refunds? Before or after month-end adjustments? The agent picked one. It didn’t say which, and it likely didn’t know there was a choice.</p><p>That question, more than any model benchmark, tells you whether an agentic data project will hold up in a real business.</p><h3>The database was never meant to be read like this</h3><p>Enterprise databases are built to be operated, not understood. Table names reflect decisions made years ago. Columns like amt, amt_adj, and amt_final_v2 sit side by side. A &quot;customer&quot; might live in several tables, each with its own idea of what counts as active.</p><p>Experienced analysts work through this with institutional knowledge. They know which flag excludes test transactions, which join double-counts jointly owned accounts, and which revenue figure the CFO actually reports.</p><p>An agent pointed at a raw schema has none of that context. It reads column names and makes reasonable guesses, and reasonable guesses are the problem, because they look right.</p><h3>The evidence is not subtle</h3><p>This isn’t just intuition. I asked Claude to check and found that Researchers at data.world built a benchmark on an insurance-industry schema and tested how well an LLM answered business questions. Using GPT-4 with zero-shot prompts directly on the SQL database, accuracy was 16%; it rose to 54% when the same questions were posed over a knowledge graph representation of that database. The gap was widest on complex schemas, where questions against high schema complexity scored 0% accuracy without the knowledge graph.</p><p>That study used a 2023 model, and today’s models are more capable. But the underlying lesson has held: models do far better when business meaning is made explicit than when they have to infer it from table structures.</p><h3>Agents remove the friction that used to catch mistakes</h3><p>When an analyst writes a bad query, something usually gives it away. It runs too long, the numbers feel off, or a reviewer catches it. Agents skip that friction.</p><p>A query can be syntactically valid and still be wrong. It runs, the result looks plausible, and the agent presents it with the same confidence as a correct answer. Nothing in the output says a join fanned out, a filter was missed, or “revenue” became gross bookings.</p><p>Multiply that across hundreds of users and thousands of questions, and you haven’t built self-service analytics. You’ve built a system that can produce competing versions of the truth very quickly.</p><h3>Documents raise the stakes</h3><p>Much of what an enterprise knows lives outside tables, in contracts, credit memos, policies, call transcripts, and regulatory filings. The common way to give agents access is retrieval-augmented generation: split documents into chunks, embed them, and retrieve passages that best match the question.</p><p>Done naively, that’s the unstructured version of handing an agent a raw schema.</p><p>Similarity search ranks text by resemblance. Unless you add governed metadata, it can’t tell whether a passage comes from an approved policy, a draft, a superseded version, or an email paraphrasing the policy incorrectly. All of them may look relevant.</p><p>Chunking creates its own problems. Anthropic has documented that traditional RAG systems often lose context when documents are split into smaller chunks, which leads to retrieval errors. A sentence like “this limit does not apply” means little once it’s separated from its heading, product, and effective date.</p><p>Access control is the most serious gap. OWASP’s RAG security guidance describes a failure where a user with access to one department’s documents retrieves chunks from another department’s restricted documents because the vector store has no access control boundaries. The concern is significant enough that the 2025 OWASP Top 10 for LLM Applications added Vector and Embedding Weaknesses as a new category.</p><p>A wrong number sometimes looks wrong. A confidently quoted, outdated policy sounds exactly like a correct one.</p><h3>What a semantic model gives you</h3><p>A semantic model translates between how the business talks and how the data is stored. It defines the business entities that matter and maps them to physical tables. It holds the approved calculation for each metric, including the filters and exclusions that make it correct. It records valid dimensions, correct join paths, and who is allowed to see what.</p><p>For analysts, that’s a convenience. For agents, it’s the difference between reasoning about the business and guessing about the database.</p><p>The same principle extends to documents. The customer defined for the warehouse becomes the same customer linked in contracts and correspondence, often through a knowledge graph. Documents carry governed metadata such as type, owner, status, effective dates, and classification. Key facts like renewal dates, contract values, and approval limits are extracted into structured form, each linked back to its source passage. And permissions from the source system travel with the content into every chunk and every extracted fact.</p><p>This is also why a built-in semantics and governance layer sits at the center of how we approach agentic AI at TechX, particularly for financial services, where a wrong answer carries regulatory consequences as well as business ones.</p><h3>This is where the industry is heading</h3><p>Semantic layers aren’t a niche idea. Tools such as dbt’s MetricFlow, Cube, and LookML have been doing this for BI for years, and they’re now being exposed directly to agents. dbt’s MCP server, for example, gives AI systems tools like list_metrics, get_dimensions, and query_metrics so they can query validated business metrics instead of writing SQL from scratch.</p><p>The industry is also working on portability. Snowflake, Salesforce, dbt Labs, and others launched the Open Semantic Interchange, which has moved to the Apache Foundation as Apache Ossie (Incubating). Its purpose is to address the problem of the same KPI being defined differently across tools, so business meaning can be defined once and used by dashboards and agents alike.</p><p>When competing vendors cooperate on a shared standard for business meaning, that’s a strong signal about where agent architecture is going.</p><h3>Semantics first, retrieval second</h3><p>The architecture isn’t complicated. The discipline to keep it is the hard part.</p><p>By default, agents don’t see the raw schema or the raw index. They get tools that expose the semantic model: list metrics, describe a metric, list valid dimensions, and find approved documents for an entity. When an agent asks for recognized revenue by region for Q2, the semantic layer generates the SQL from the approved definition. When it asks about the current refund policy, retrieval filters for approved, in-effect documents the user is permitted to see, and only then ranks by relevance.</p><p>The agent’s job shifts from writing queries to choosing the right business concepts.</p><p>Ambiguity gets handled differently too. If someone asks for “revenue” and the model defines three kinds, the agent presents the options and their definitions and asks which one is meant.</p><p>Access control is enforced below the agent, not described in the prompt. OWASP recommends query-time filtering that enforces the querying entity’s access boundaries before similarity search results are returned. An agent can’t expose data it was never allowed to retrieve.</p><p>Raw access still has a place for exploratory work. It should be the exception: explicit, logged, limited to specific roles, and clearly labeled as ungoverned.</p><h3>The valuable questions need both</h3><p>The strongest case for one semantic model across structured and unstructured data is that high-value questions span both.</p><p>Consider: “Which of our top customers by revenue have contracts renewing next quarter with pricing below our current standard rate?”</p><p>Revenue ranking comes from the warehouse. Renewal dates and pricing terms come from contracts. The standard rate comes from a pricing policy. An agent working over a raw schema and a raw index has to guess how a table row relates to a PDF. An agent working over a semantic model follows a defined path, from customer entity to governed revenue metric to linked contracts to extracted terms to the current approved policy, and can cite each step.</p><h3>Common objections</h3><p><em>Building a semantic model takes too long.</em> You don’t need to model the entire enterprise first. Start with the metrics that drive high-stakes decisions and the documents people rely on most, then expand based on real demand.</p><p><em>Agents need flexibility.</em> Flexibility without correctness is a liability. Let the governed path handle most questions and route the rest through a controlled exception path.</p><p><em>The model will never be complete.</em> That’s useful. When an agent can’t answer from governed sources, log the gap. Those logs become a backlog grounded in what people actually asked.</p><p><em>Can’t we put the definitions in the prompt?</em> That works in a demo. Prompts can be versioned, but they aren’t enforced as policy and aren’t shared across every agent and dashboard in the organization. A semantic model is built to be all of those things.</p><h3>Questions worth asking this week</h3><p>Can your agents query the raw warehouse or search an ungoverned document index today, and if so, who approved that and is it logged? Does each top metric have one approved definition? Do important documents carry status and effective dates, and does retrieval filter on them? Are core entities like customer and product linked consistently across databases and documents? Are access rules enforced at retrieval time, or only described in agent instructions? Can every answer point to the metric definition or source passage behind it? And do you track the questions agents couldn’t answer?</p><p>If most answers are “no” or “not sure,” the agent isn’t the first thing to fix.</p><h3>The bottom line</h3><p>Connecting an agent straight to your data is tempting. It’s fast, it demos well, and it feels like progress. But an agent reasoning over raw tables and raw text is only as trustworthy as its guesses, and guesses about money, customers, and risk aren’t good enough.</p><p>Put the semantic model between your agents and all of your data, tables and text alike. Make business meaning the interface, and let SQL and search be implementation details. Your numbers and your documents will start agreeing with each other, and the next time someone asks “which revenue?” or “which policy?”, the system will already know.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5071a0983ac3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[At TechX, We’re Treating Agentic Workflows as a Product. Here’s What Broke, and How ADLC Fixed It.]]></title>
            <link>https://medium.com/@alexpongpech/at-techx-were-treating-agentic-workflows-as-a-product-here-s-what-broke-and-how-adlc-fixed-it-4310d7043c29?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/4310d7043c29</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 01:08:42 GMT</pubDate>
            <atom:updated>2026-09-17T01:08:42.839Z</atom:updated>
            <content:encoded><![CDATA[<p>A demo can make almost any agent look good. You ask the right question, it calls a tool, returns a clean answer, and the room nods. Then you ship it, and real people use it in ways nobody planned for.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*F4bk4-B1tgcNRfqm" /><figcaption>Photo by <a href="https://unsplash.com/@charlesdeluvio?utm_source=medium&amp;utm_medium=referral">charlesdeluvio</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>At TechX, we build agentic workflows as a product for the financial industry. People rely on them, and regulators may look closely. That meant we couldn’t treat agents as clever automation. We needed a lifecycle built for them: the Agent Development Lifecycle, or ADLC.</p><p>We didn’t start there. We got there by running into problems. Here are the ones that shaped how we work.</p><h3>Problem 1: A small prompt edit quietly broke things</h3><p>A wording change meant to improve tone changed how the agent handled a certain request. Nothing in the code changed, so nothing in our pipeline caught it. The agent just started behaving differently.</p><p><strong>How ADLC helped:</strong> We stopped treating prompts as configuration someone could tweak on the side. Prompts, context, tool permissions, and policies are now <strong>versioned release artifacts.</strong> Any change goes through review and evaluation, just like code.</p><h3>Problem 2: Every test passed, and the agent was still wrong</h3><p>Our tests checked whether the system worked. They didn’t check whether the agent’s answers were right. Ask the same question twice, and it might take a different path, sometimes a bad one.</p><p><strong>How ADLC helped:</strong> We built evaluation suites that run each agent across many realistic scenarios and measure the results together. How often is it accurate? Does it stick to its sources? Does it stay within policy? Each release has to clear a set bar. If it doesn’t, it waits, even if the demo looks great.</p><h3>Problem 3: The agent sounded confident but used the wrong information</h3><p>We encounter this problem often when we developed our internal KM agentic workflow. The reasoning was fine. The context was the problem. Business data changes constantly, and the agent had no way of knowing what was stale.</p><p><strong>How ADLC helped:</strong> We manage context and semantics with its own lifecycle. We track sources, test how well answers are grounded, and review what the agent can retrieve. Grounding is now one of the core measures in every evaluation.</p><h3>Problem 4: A model update changed behavior overnight</h3><p>We didn’t touch anything. The underlying model changed, and some workflows shifted with it.</p><p><strong>How ADLC helped:</strong> We <strong>pin model versions</strong>. Before any upgrade, we run the full evaluation suite against the new model and compare results. Upgrades are planned releases, not surprises.</p><h3>Problem 5: We couldn’t explain what the agent had done</h3><p>When someone asked why an agent took a certain action, piecing the answer together took far too long. In a regulated industry, that isn’t acceptable.</p><p><strong>How ADLC helped:</strong> <strong>Traceability</strong> is now a release requirement. For any action, we can see what the agent saw, what it decided, which tools it called, and which version of its instructions it followed. If we can’t explain it, it doesn’t ship. We built this into our governance engine.</p><h3>Problem 6: Autonomy moved faster than trust</h3><p>Early on, it was tempting to let agents do as much as possible. Some of the steps they could take had real consequences, specifically in the financial industry.</p><p><strong>How ADLC helped:</strong> Every workflow step has a risk tier. Low-risk steps run on their own. Higher-risk steps stay behind human approval until evaluation data justifies loosening that. Agents get only the tool access they need. Adding autonomy over time is much easier than taking it back after something breaks.</p><h3>Problem 7: Production failures didn’t make us better</h3><p>A problem would come up, we’d fix it, and a similar one would show up later.</p><p><strong>How ADLC helped:</strong> Launch isn’t the finish line. We monitor agents in production, and every real failure goes into our evaluation set as a new test case. Over time, the suite reflects what actually goes wrong, not just what we predicted.</p><h3>What we’ve learned</h3><p>None of these problems were about the model not being smart enough. They came from treating<strong> a probabilistic system</strong> like ordinary software.</p><p>ADLC didn’t replace our engineering discipline. It extended it to cover what we actually ship: code, prompts, context, tools, and the evidence that the agent is behaving. We haven’t figured everything out. There’s no agreed ADLC standard yet, and we still debate how much evaluation is enough and where people should stay in the loop.</p><p>But one thing is clear. A demo shows what an agent can do. The lifecycle shows what it will do, day after day, when nobody’s watching. At TechX, that’s what we’re building for.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4310d7043c29" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[It’s Not My First Rodeo, But It Is Always My First]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@alexpongpech/its-not-my-first-rodeo-but-it-is-always-my-first-1bfc54815119?source=rss-6ec1802f3d5c------2"><img src="https://cdn-images-1.medium.com/max/2600/1*AggNhUlhSWTnIQIKsU1bFQ@2x.jpeg" width="3240"></a></p><p class="medium-feed-link"><a href="https://medium.com/@alexpongpech/its-not-my-first-rodeo-but-it-is-always-my-first-1bfc54815119?source=rss-6ec1802f3d5c------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@alexpongpech/its-not-my-first-rodeo-but-it-is-always-my-first-1bfc54815119?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/1bfc54815119</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 10:24:32 GMT</pubDate>
            <atom:updated>2026-09-01T10:24:32.367Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Exploring Agentic Data Engineering: Governance, Infrastructure, and the Knowledge Layer]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@alexpongpech/exploring-agentic-data-engineering-governance-infrastructure-and-the-knowledge-layer-e6682968b058?source=rss-6ec1802f3d5c------2"><img src="https://cdn-images-1.medium.com/max/2160/1*tpR5PEPyuQJ1WRjw82e_8w@2x.jpeg" width="2160"></a></p><p class="medium-feed-link"><a href="https://medium.com/@alexpongpech/exploring-agentic-data-engineering-governance-infrastructure-and-the-knowledge-layer-e6682968b058?source=rss-6ec1802f3d5c------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@alexpongpech/exploring-agentic-data-engineering-governance-infrastructure-and-the-knowledge-layer-e6682968b058?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/e6682968b058</guid>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Sat, 23 May 2026 02:51:45 GMT</pubDate>
            <atom:updated>2026-05-23T02:51:45.972Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Governance Bottleneck: Why Your Agentic AI Chain Is Only as Strong as Its Weakest Constraint]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-link"><a href="https://medium.com/@alexpongpech/the-governance-bottleneck-why-your-agentic-ai-chain-is-only-as-strong-as-its-weakest-constraint-e34a85e67c2b?source=rss-6ec1802f3d5c------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@alexpongpech/the-governance-bottleneck-why-your-agentic-ai-chain-is-only-as-strong-as-its-weakest-constraint-e34a85e67c2b?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/e34a85e67c2b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Sat, 23 May 2026 02:00:09 GMT</pubDate>
            <atom:updated>2026-05-23T02:52:09.612Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Over-Engineering Trap: Building an Agentic AI Platform for Value, Not Vanity]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@alexpongpech/the-over-engineering-trap-building-an-agentic-ai-platform-for-value-not-vanity-8bff4370390a?source=rss-6ec1802f3d5c------2"><img src="https://cdn-images-1.medium.com/max/2160/1*_psAylGQsFzoVC3NNKOUEg@2x.jpeg" width="2160"></a></p><p class="medium-feed-link"><a href="https://medium.com/@alexpongpech/the-over-engineering-trap-building-an-agentic-ai-platform-for-value-not-vanity-8bff4370390a?source=rss-6ec1802f3d5c------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@alexpongpech/the-over-engineering-trap-building-an-agentic-ai-platform-for-value-not-vanity-8bff4370390a?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/8bff4370390a</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[leadership]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Wed, 06 May 2026 05:34:12 GMT</pubDate>
            <atom:updated>2026-05-06T05:34:12.425Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[The Distributed Blueprint for Agentic AI: Why AMS Must Think Like a Database]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@alexpongpech/the-distributed-blueprint-for-agentic-ai-why-ams-must-think-like-a-database-74e1a9630c9f?source=rss-6ec1802f3d5c------2"><img src="https://cdn-images-1.medium.com/max/2160/1*MWvCxVH_sAb72-EpOBYUYQ@2x.jpeg" width="2160"></a></p><p class="medium-feed-link"><a href="https://medium.com/@alexpongpech/the-distributed-blueprint-for-agentic-ai-why-ams-must-think-like-a-database-74e1a9630c9f?source=rss-6ec1802f3d5c------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@alexpongpech/the-distributed-blueprint-for-agentic-ai-why-ams-must-think-like-a-database-74e1a9630c9f?source=rss-6ec1802f3d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/74e1a9630c9f</guid>
            <category><![CDATA[distributed-systems]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Alex Worapol Pongpech]]></dc:creator>
            <pubDate>Sun, 03 May 2026 10:11:54 GMT</pubDate>
            <atom:updated>2026-05-03T10:11:54.489Z</atom:updated>
        </item>
    </channel>
</rss>