<?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 Mel Meng on Medium]]></title>
        <description><![CDATA[Stories by Mel Meng on Medium]]></description>
        <link>https://medium.com/@mel-meng-pe?source=rss-1602561a2c72------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*f0CPLA80a_S5kai3QQZpLw.jpeg</url>
            <title>Stories by Mel Meng on Medium</title>
            <link>https://medium.com/@mel-meng-pe?source=rss-1602561a2c72------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 16:06:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mel-meng-pe/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[Trust Is What Makes a Model a Model]]></title>
            <link>https://mel-meng-pe.medium.com/trust-is-what-makes-a-model-a-model-beaab9f90d37?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/beaab9f90d37</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 14:08:44 GMT</pubDate>
            <atom:updated>2026-10-15T14:08:44.206Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cAT4JOHntaLhFlQRTpfocA.png" /></figure><p>Subaru has a well-known tagline: “Love. It’s what makes a Subaru, a Subaru.”</p><p>Notice what it leaves out. It doesn’t mention the engine, the drivetrain, or the specifications. Those matter, but they aren’t what the brand says it is.</p><p>Models are similar.</p><p>A model built for a utility isn’t really about the pipes, the simulation engine, or any of the technical details it’s made of. It’s about whether a utility can confidently make decisions involving millions of dollars that ultimately affect people’s lives.</p><p>Imagine you’re reviewing a stormwater model for a proposed subdivision. You don’t care how many pipes it contains or how sophisticated the software is. You want to know:</p><ul><li>Will the development cause flooding or make downstream flooding worse?</li><li>Does the proposed drainage system have enough capacity?</li><li>Can you trust the model results?</li></ul><p>To answer these, you need to understand what the model is supposed to do, what assumptions were made, whether the simulation runs correctly, and whether the results make sense.</p><p><strong>That’s what a model checkup is.</strong></p><h3>Build the trust, not just the model, from day one</h3><p>Now switch roles. You’re the engineer building the model.</p><p>It’s easy to think the job is to build the model and then convince people it’s right. But trust isn’t something you add at the end. You build it alongside the model.</p><p>Keep an imaginary reviewer in your head, someone who will eventually ask those three questions. Every time you run the model, ask what they would ask.</p><p>That’s why I recommend running your model as early as possible, ideally on the first day, even if it’s one pipe and one design storm. A first run is where you find the wrong units, the disconnected node, or the missing outfall, while they’re still easy to fix.</p><p>Then keep checking as you build.</p><p>Every checkup is a chance to catch a mistake, challenge an assumption, and learn something about the system you’re modeling. By the time a reviewer sees the model, you’ve already asked their questions many times.</p><p><strong>Model checkup isn’t something you do after building a model. It’s how you build a better model.</strong></p><h3>The Model Checkup tool: building trust into the modeling process</h3><p>Most engineers already check their models. But much of that work happens outside the model, in spreadsheets, reports, or sometimes just in our heads.</p><p><strong>A model report tells people what the model predicts. A documented checkup helps them understand why you trust those predictions.</strong></p><p>What if checking and documenting our work were as much a part of modeling as adding another pipe?</p><p>That’s the idea behind the interactive <strong>Model Checkup tool</strong> I created. It’s where the answers to the reviewer’s questions get written down.</p><p>It organizes the process into five areas:</p><p><strong>Project → Model → Simulation → Findings → Troubleshooting</strong></p><p>One distinction I find particularly useful is separating results, findings, and troubleshooting.</p><p>For example, suppose residents have reported flooding at an intersection, but your model predicts none.</p><ul><li><strong>Result:</strong> The model predicts no flooding at that intersection.</li><li><strong>Finding:</strong> The prediction doesn’t appear consistent with the reported flooding.</li><li><strong>Troubleshooting:</strong> Was the reported flooding caused by a larger storm? Are the pipe elevations correct? Is a drainage structure missing?</li></ul><p><strong>The goal isn’t to make the model match your expectations. It’s to understand why the results are different, and to document what you’ve learned.</strong></p><p>That’s how trust gets built: not by claiming the model is right, but by showing you’ve asked why it might not be.</p><p>I’ve included sample projects to illustrate the process. You can explore those or use the tool to document a checkup of your own model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2wE1pSD8Z87hTHqKDLVCiw.png" /></figure><p><a href="https://mel-meng.github.io/model-checkup/">Explore the Model Checkup tool</a></p><p>I hope you find it useful. I’d love to hear how you approach model checkup and documentation in your own projects.</p><p><em>Trust. It’s what makes a model a model.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=beaab9f90d37" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[From Clicks to Decisions: How AI Is Changing My Software Training Class]]></title>
            <link>https://mel-meng-pe.medium.com/from-clicks-to-decisions-how-ai-is-changing-my-software-training-class-b52f33136910?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/b52f33136910</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Mon, 12 Oct 2026 18:35:59 GMT</pubDate>
            <atom:updated>2026-10-12T18:35:59.496Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EBgGA8UxR3D6B3NktHid4A.png" /></figure><p>I love modeling. I might love teaching it even more.</p><p>But I don’t miss the countless nights fixing models before a morning deadline. I learned a lot the hard way: building models that didn’t work, and tracing bad results back to the choices I’d made. I’ve always wished I could help others get there without all that frustration.</p><p>Something has always bothered me about software training, though. We design careful exercises, write detailed instructions, and walk students through every step to a perfect result. By then, we’ve made all the decisions for them. Where’s the fun in that?</p><p>We do it for a reason. The software is hard, and students can’t make any decisions until they can use it.</p><p>Recently, I hit a strange problem in QGIS on my 4K monitor: the 3D view wouldn’t fill the window. A quick Google AI search pointed me in the right direction, and a few minutes later I was back to what I actually wanted to do.</p><p>That got me thinking. Without AI, I couldn’t skip teaching the basics. A student stuck on the software is stuck on everything, so someone has to walk them through every step. Now they can just ask AI. That means I can start with the decisions instead of the clicks.</p><p>So in my latest water modeling course, I’m trying something different.</p><p>Instead of instructions for one perfect mesh, I ask students to build five or six different meshes for the same cul-de-sac.</p><p>Change the parameters. Compare the results.</p><p>The point isn’t to make six meshes. It’s to see what happens when you make different choices. Which details matter? What are the trade-offs? Which mesh would you actually use, and why?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hx5hRjW0x0XoNRXQaBiRVQ.png" /></figure><p><em>Same cul-de-sac, different meshes. Which one would you choose, and why?</em></p><p>I think that’s how we develop judgment. Not by following instructions to the right answer, but by exploring different possibilities, including the wrong ones, and learning from what happens. My job is to sit with them in those choices and ask the questions that make them think.</p><p>That’s how I actually learned. I got better by trying things, seeing unexpected results, and figuring out why. Over time, I developed a feel for when something was off, sometimes before I could explain it.</p><p>I’d like my students to have that experience, without the frustration that comes from fighting the software instead of the problem.</p><p>Much of the software training I’ve seen uses AI to teach the same steps more efficiently. For me, it’s different. AI let me choose what not to teach. It can answer a student’s software question at 11 p.m., in their own words, as many times as they need. I can’t. So I can spend our time on what only I can offer.</p><p>I don’t know yet how well this experiment will work. But I’m excited about spending less time showing students where to click and more time watching them wrestle with the decisions.</p><p>After all, that’s the part of modeling, and teaching, that I’ve always loved.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b52f33136910" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Good Model Is a Thin Book]]></title>
            <link>https://mel-meng-pe.medium.com/a-good-model-is-a-thin-book-99e7480af59a?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/99e7480af59a</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 14:44:15 GMT</pubDate>
            <atom:updated>2026-10-02T14:44:15.052Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JFYTNPuqPrammG9H6cb32Q.png" /></figure><p>There is a Chinese saying about learning that I have always liked: first read the book until it becomes thick, then read it until it becomes thin again.</p><p>The thin book is where learning ends up. It isn’t the book of someone who knows less. It is the book of someone who has compressed everything into what matters.</p><p>Hydraulic modeling is the same. A good model is the thin book.</p><h3>What I used to believe</h3><p>When I started building fine-resolution 2D models at the watershed and city scale, I thought better technology would reduce how much judgment I needed. Better terrain data, finer meshes, faster computers. Put enough reality into the model, and the physics could sort out the rest.</p><p>With a GIS background, this came naturally. More data and more scripting meant more detail, and I hoped those fun, simple data conversion tasks would fix my downstream headaches: complicated flow patterns and long run times.</p><p>The opposite turned out to be true. Every mesh element costs computation, and every object is one more thing to build, check, and get wrong. Push the detail far enough and runs take forever, debugging gets harder, and strange results appear everywhere.</p><p>At that scale, simplification is not a preference. It is a constraint. So the real question is what to keep.</p><h3>Model the function, not the detail</h3><p>I don’t care about the exact dimensions of a curb inlet. I care whether it is supposed to capture most of the flow on the street. I don’t care whether a roof is flat or steeply sloped. I care whether it drains to the street or onto the lawn, because that changes how quickly runoff reaches the street.</p><h3>Compression is the destination</h3><p>You can only decide what to remove once you understand how the system works. Where should the water go? What controls that path? What keeps it from going elsewhere? That messy reality is the thick book, and you can’t skip it.</p><p>It also makes the model more useful. When results disagree with you, the disagreement is information, not a surprise. Maybe your understanding was wrong, the data is wrong, or the mesh lost an important feature.</p><p>Simplification without understanding is information loss. Simplification with understanding is compression. Compression is where learning ends up: not more detail, but a real understanding of the system.</p><p>A good model doesn’t mimic how a watershed looks. It learns how the watershed thinks: where the water wants to go, and what decides whether it gets there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=99e7480af59a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Might Be the Best Coach I’ve Had for Understanding SWMM5 Runoff]]></title>
            <link>https://mel-meng-pe.medium.com/ai-might-be-the-best-coach-ive-had-for-understanding-swmm5-runoff-5cd0fb5ae586?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/5cd0fb5ae586</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Wed, 30 Sep 2026 00:51:14 GMT</pubDate>
            <atom:updated>2026-09-30T00:51:14.029Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dOAomMqy2g2oA2hHuN5-4A.png" /></figure><p>I’ve always found the SWMM5 runoff method surprisingly hard to explain.</p><p>Not because the individual ideas are that complicated. Rainfall, infiltration, depression storage, surface runoff — each one is fairly simple.</p><p>The hard part is seeing how they all work together.</p><p>I’ve wanted for years to find a better way to explain that, so recently I started building a small teaching tool around the SWMM5 runoff method.</p><p>AI wrote essentially all of the code. And it did some genuinely amazing work. Sometimes it built things much faster than I could have. Sometimes it found solutions I would not have thought of myself.</p><p>What surprised me was how that made me feel. I never once felt threatened by how capable it was.</p><p>I just wanted it to be even better.</p><p>Then I realized why.</p><p>It felt more like working with a great coach.</p><p>If your goal is to get better at something, you don’t want a coach who only sees what you already see. You want someone who notices things you miss, gives you another way to look at the problem, and pushes you further.</p><p>The better the coach, the better.</p><p>That is how AI felt on this project.</p><p>I still had to decide what the tool should teach. I still had to look at the behavior and say, “No, that isn’t how runoff works,” or, “This still doesn’t make the idea clear enough.”</p><p>But AI gave me a way to keep trying ideas until I could finally build something close to what I had imagined for years.</p><p>And that is also what I hope the tool can do for you learning SWMM.</p><p>Not do the modeling for you, but help you understand by playing with the toy tools.</p><p><strong>Try it here:</strong> <a href="https://mel-meng.github.io/swmm-labs/">SWMM Labs</a></p><p><em>Thanks to Robert Dickinson and the open-source </em><a href="https://github.com/SWMMEnablement/SWMM-Engine"><em>SWMM-Engine</em></a><em> project, which makes it possible to run the real SWMM5 engine directly in the browser through WebAssembly.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5cd0fb5ae586" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Hydraulic Modeling Can Learn from Car Manufacturing]]></title>
            <link>https://mel-meng-pe.medium.com/what-hydraulic-modeling-can-learn-from-car-manufacturing-afaae36719c1?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/afaae36719c1</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 21:15:18 GMT</pubDate>
            <atom:updated>2026-09-25T21:15:18.103Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Use Data Flags to Keep Model-Building Work Moving</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0rjamqKMokOzgVVHyMKWDA.png" /></figure><p>When a car is built, thousands of parts come from many different suppliers.</p><p>Those parts do not go straight from the delivery truck into the finished vehicle. They move through a production process, and some require inspection or rework before they are ready to install.</p><p>At Toyota, parts move with Kanban cards, small cards that travel with the work and help signal what is needed and what happens next.</p><p>Hydraulic model data is surprisingly similar.</p><p>We may start with GIS, LiDAR, plans, existing models, survey, field records, and other sources. Some information is ready to use. Some needs review. Some is missing. Some must be inferred or temporarily assumed. And some source data needs to be processed before it becomes useful in the model.</p><p>The challenge is not simply getting better data. It is <strong>keeping the model-building work moving without losing track of where each piece of data stands</strong>.</p><h3>Where we often get data flags wrong</h3><p>Data flags are often treated as a one-time data-quality stamp. A value came from GIS, so we flag it GIS. A value came from survey, so we flag it survey. A value was assumed, so we flag it assumed. Done.</p><p>But a flag can do much more than document where a number came from. It can tell us <strong>where that piece of data is in the workflow and what should happen next</strong>.</p><p>If a pipe invert is flagged INF(inferred), for example, we immediately know the original data was missing or unreliable, the value was inferred so modeling could continue, and better information may still be needed.</p><p>But not every inferred invert deserves a survey crew. Whether this one does depends on what the model tells us about that location.</p><p>Instead of treating flags as stamps applied to finished data, we can use them like Kanban cards for model building: make the current state visible, keep the work moving, and show what still needs attention.</p><p>A source flag tells us where the information came from. A workflow flag tells us <strong>where it is in the model-building process and what happened to it</strong>. Together, they let us work with imperfect data without losing control of the model-building process.</p><h3>The other half</h3><p>Kanban cards do more than travel with the parts. At Toyota, they also help signal when parts should move.</p><p>That turns out to be the more interesting idea for hydraulic modeling. Once we can see which parts of the model are still uncertain, the next question is <strong>which ones deserve more work first</strong>.</p><p>That is a topic for the next article.</p><p>For a practical one-page reference, I put together a cheat sheet based on the City of Houston’s InfoWorks ICM flagging conventions:</p><p><a href="https://mel-meng.github.io/houston/houston_icm_data_flag_cheatsheet.html">Houston-Style ICM Data Flag Cheat Sheet</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=afaae36719c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Lessons from a Model Migration: Learning by Building the Wrong Tool]]></title>
            <link>https://mel-meng-pe.medium.com/what-the-engineering-loop-looks-like-a-personal-story-12918502e427?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/12918502e427</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 14:50:25 GMT</pubDate>
            <atom:updated>2026-09-16T15:12:33.801Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cS-sqjFDYHUq-CU0YrwtbQ.png" /></figure><p>A customer was migrating a model from older software. He wasn’t after new capability; he wanted the new model to behave like the old one.</p><p>Sounds simple. I also knew how painful this kind of project could be.</p><p>A few years ago, I would have converted what I could, fixed the rest manually, exported a pile of tables, compared results, chased the differences, and explained them one by one. It would work. It would also be tedious.</p><p>This time, coding agents had gotten good enough that I got a little more ambitious: what if, instead of doing all this checking myself, I built something the customer could actually use?</p><p>The first part went surprisingly well. With AI helping me code, I closed a few gaps in the converter within a few hours. That gave me confidence — probably too much.</p><p>I started thinking: why not build a comparison tool that checks everything? Every pipe, every node, every time step. The customer could do the tedious comparison instead of me, every difference, all at once.</p><p>What a brilliant idea.</p><p>By the end of the first day, I realized I had built the wrong thing. There was simply too much information. I didn’t want to open it myself. How was I going to convince the customer to?</p><p>So I went back with a simpler question: <strong>What do you actually look at when deciding whether a model makes sense?</strong></p><p>Peak flows. How full the pipes get. A handful of key results.</p><p>It took about half a day to build that version.</p><p>Then we compared the results.</p><p><strong>They didn’t match.</strong></p><p>That was the part I really didn’t want to see.</p><p>After a few hours slicing and rotating the results across tables and maps, I narrowed the problem down to a handful of nodes. The loadings were zero there in the old model. AI helped me with the last step: reviewing the log file turned up clashing IDs, which explained why those loadings had been dropped.</p><p>I showed the customer what was different, where it came from, and why. I was braced for the worst. Was migration still an option after this?</p><p>To my surprise, he was happy we had found it.</p><p>And we moved on.</p><p>Neither of us had ever said that every result had to match exactly, but that was the assumption we started with. Once we had enough evidence, we both knew the goalpost had moved. Nobody announced it.</p><p>The real acceptance criteria weren’t written down anywhere. We discovered them by doing the work.</p><p>What we had actually built over those few weeks was trust, in both directions. He learned that I would keep digging until I understood a difference. I learned which differences actually mattered to him.</p><p>Looking back, Ray Dalio’s formula felt a little less abstract:</p><p><strong>Pain + reflection = progress.</strong></p><p>The pain in projects like this is easy to find. Reflection has always been harder for me. After a consulting project, there are dozens of things I <em>could</em> learn. Which ones are worth carrying forward?</p><p>Bringing AI coding into the work had a surprising side effect: <strong>I started treating the job more like a product problem.</strong></p><p>The question changed from <em>How do I get through this?</em> to <em>How do I build something the customer could use without me?</em></p><p>Neither tool ended up being the deliverable. The first one was thrown away after a day. But building for the customer forced me to look at his problem instead of just my workload.</p><p>It also made the reflection more concrete: <strong>If I could change the software before the next migration, what would I change?</strong></p><p>For this project, I already have an answer:</p><ul><li>Help customers see where results differ.</li><li>Help them understand why.</li></ul><p>That’s why <a href="https://www.latent.space/p/forward-deployed-engineer-best-practices">Vinoo Ganesh’s piece on Forward Deployed Engineers</a> resonated with me. He describes a similar loop: get close enough to the customer to learn something you couldn’t have known from the outside, then let that shape what gets built next.</p><p>If this is part of the future of engineering, I like it already.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=12918502e427" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is a model for?]]></title>
            <link>https://mel-meng-pe.medium.com/what-is-a-model-for-44ee6573b35c?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/44ee6573b35c</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Sat, 12 Sep 2026 23:16:10 GMT</pubDate>
            <atom:updated>2026-09-12T23:16:10.880Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ugcpb1FsWjxtUjXF25vgsg.png" /></figure><p>I am putting together a hydraulic modeling training, and I got stuck at the first slide: <strong>What is a model for?</strong></p><p>I have been modeling since the early 2000s, so you would think I should have a good answer. Early in my career, I probably would have said the job was to build the model correctly: right equations, right data, calibrated, checked.</p><p>But eventually the questions get harder. A city might ask: <em>If we spend $20 million over the next 10 years, the flooding problem goes away, right? Your model says so. You’re confident, right?</em></p><p>Am I? Is the model “right”?</p><p>Not the way 2 + 2 = 4 is right. We are predicting rainfall, development, and infrastructure years into the future. We can always collect more data, calibrate one more thing, or add more detail. There is no clean point where the model suddenly becomes correct.</p><p>The textbook answer is that <strong>a model is a decision tool</strong>. That is true, but I think the modeler’s job goes one step further: help people understand <strong>how much confidence the result deserves</strong>. Not “trust me,” but: <em>I’m confident about this. I’m less confident about that, but it probably won’t change the decision. This assumption might, so we should look harder.</em></p><p>Recently, using AI coding agents put me on the other side of that same problem. An agent works for a few minutes and comes back: <strong>Done.</strong> And now I’m asking exactly what my clients have asked me for years: <em>How do I know? How much should I trust this?</em></p><p>Maybe that is one thing great engineers learn to do well. Not just produce the work, but explain <strong>how much confidence the work deserves, and why.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=44ee6573b35c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pick Your Own Hard]]></title>
            <link>https://mel-meng-pe.medium.com/pick-your-own-hard-c11eba72151e?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/c11eba72151e</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Fri, 04 Sep 2026 00:51:38 GMT</pubDate>
            <atom:updated>2026-09-04T01:13:03.580Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vfxT3Tt72d22LAH3H4WO7w.png" /></figure><p>The more I use AI, the more I feel the boundary moving. Things I spent years learning can suddenly be done in minutes. And every so often, that creates an uncomfortable question:</p><p><strong>If AI gives me an easier path, should I always take it?</strong></p><p>My gut tells me I’ll miss those <em>aha</em> moments that come after struggling with something for a while. You’re confused, you try a few things, and then eventually: <strong>Oh. Now I get it.</strong></p><p>I enjoy that process of learning something hard. But lately I’ve realized there’s something deeper in it for me. When I understand something that way, almost immediately I want someone else to experience it too.</p><p>That has probably been true since I was a kid helping other kids with math. Years later, when I joined Innovyze, the phrase <strong>“Empowering Water Experts”</strong> resonated with me immediately. I didn’t think much about why at the time.</p><p>Using AI has made the reason much clearer to me. Sharing what I understand — not just the answer, but the understanding itself — means a lot to me.</p><p>That’s also why Terence Tao’s <a href="https://teorth.github.io/tao-web/slides/age-of-ai-icm-2026.pdf">thoughts</a> about AI and mathematics stuck with me. What resonated was the idea that mathematics isn’t only about solving problems. It is also about turning what we learn into knowledge that helps more people do math.</p><p>I think I want the same thing from engineering. AI can remove a lot of struggle, and I’m happy to let it. But I’m becoming more careful about which struggle I give away.</p><p><strong>Not because hard is better, but because earned understanding is.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c11eba72151e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Attention Is All You Need in the Age of AI]]></title>
            <link>https://mel-meng-pe.medium.com/attention-is-all-you-need-in-the-age-of-ai-a4925497f38c?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/a4925497f38c</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Thu, 03 Sep 2026 00:10:52 GMT</pubDate>
            <atom:updated>2026-09-03T00:17:30.044Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*97iO1Hr_vqGgQbY90oXW1A.png" /></figure><p>When I first started as an engineer, I spent some time inspecting houses for stormwater connections to the sanitary sewer without entering the property. It was boring work: check the downspouts, look at the basement, fill out the form. Then I worked with a coworker who flipped houses on the side. Before I finished the first page of my form, he would say, “This one has a downspout connected to the basement.”</p><p>“How do you know that so quickly?”</p><p>He’d point to the roof. “This style was common in houses built in the ’80s. They usually connected the downspouts this way. And see that basement window behind the bushes?”</p><p><strong>We were looking at the same house. I had a form full of data. He had an answer.</strong></p><p>A few years later, I worked on an asset management project for a sewer utility. We wanted to know which pipes were the most critical. I knew almost nothing about asset management at that time. So we put experienced people in a room and asked embarrassingly simple questions:</p><p>How important is a school nearby? A railroad? A hospital? How deep is the pipe? Rate each from 1 to 5. Add some weights. Calculate a score.</p><p>I remember thinking: <strong>You must be kidding me. This is the model?</strong></p><p>But it worked surprisingly well. <strong>The model wasn’t smart. The experts were. </strong>We had simply captured what they paid attention to.</p><p>I think about those two experiences a lot now that I use AI.</p><p>I used to assume that if I wanted expert-level results, I had to become the expert — the hard way, with lots of time and sweat.</p><p>Maybe not always. For the first time, I can work in areas where I’m not the expert and still get surprisingly close to expert-level results. AI gives me access to knowledge I haven’t mastered myself. <strong>But I’m only useful in the process if I know what to ask, what matters, and what to pay attention to.</strong></p><p>One idea I took from <em>The Inner Game of Tennis</em> is that you don’t improve by consciously controlling every movement. You learn what to pay attention to and let much of the rest happen.</p><p>AI is making me wonder whether the same is true for knowledge work.</p><p>Maybe the question isn’t always: <strong>How do I do this like an expert?</strong></p><p>Maybe it’s: <strong>What would an expert pay attention to?</strong></p><p><em>Just watch the damn ball.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a4925497f38c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Back-of-the-envelope engineering: the most important engineering skill school doesn’t really teach.]]></title>
            <link>https://mel-meng-pe.medium.com/back-of-the-envelope-engineering-the-most-important-engineering-skill-school-doesnt-really-teach-666d43aa3a1a?source=rss-1602561a2c72------2</link>
            <guid isPermaLink="false">https://medium.com/p/666d43aa3a1a</guid>
            <dc:creator><![CDATA[Mel Meng]]></dc:creator>
            <pubDate>Wed, 02 Sep 2026 01:18:29 GMT</pubDate>
            <atom:updated>2026-09-04T12:46:58.302Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fkRwkQpB8PeDkyOO7nJeDg.png" /></figure><p>When I was a younger engineer, “checking my work” mostly meant… checking everything again. And I hated it. It was boring and slow.</p><p>Over time, I started doing something different. Instead of asking, <strong>“How do I check all of this?” </strong>I started asking, <strong>“What’s one quick thing I can do, from a different angle, that would tell me if something is seriously wrong?”</strong></p><p>I was reminded of this recently while comparing the same hydraulic model running on two versions of an engine. Hundreds of pipes. Thousands of 2D cells. <strong>How do you compare all of that without looking at every result?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nLpkj7xTI7sTa7mvvokDRA.png" /></figure><p>A simple diff of the logs did the trick. I don’t even need to read every number. One glance tells me most of the differences are down in the decimal places.</p><p>That’s what I’ve come to think of as <strong>back-of-the-envelope engineering</strong>.</p><p>Don’t do the whole thing twice. Find a quick sanity check that tells you whether you should be worried.</p><p>And strangely, after doing this for years, the checking part has become kind of fun. When the check is not obvious, it becomes a little puzzle.</p><p>Now AI can get work used to take me hours done in minutes, the back-of-the-envelope engineering I used to do at the end of the hour is <strong>becoming more and more of the job.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=666d43aa3a1a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>