Solving a $30K Engineering Problem With AI and a Textbook

A couple of years ago I was sitting at my desk, working on a pipe stress analysis for a large EV component manufacturing factory. My scope was the hot oil lines. Partway through, a question came up about the pipe shoes, the supports that bolt those very hot lines to structural steel. The concern was heat conducting through the shoe into the steel above which supported the mezzanine floor. Somebody walking by could put a hand on that steel. This was a personnel protection question, and it landed on me.
The typical operation is straightforward. You hire a specialist. A CFD expert, a thermal heat transfer person, who models the specific interface and supports and tells you whether you need an insulation barrier between the shoe and the structure. That study typically runs $20,000 to $30,000+ and takes several weeks to complete.
Our project couldn’t absorb the standard way study path in this case. I had the pressure to get the job done right, and I had less than two weeks to come up with a solution.
How do you solve a $30,000 problem with no budget and half the time?
Here’s the part I have to admit first. At that point I’d been poking at ChatGPT for about a year and a half, and I wasn’t good at it. I used it kind of like Google, and the results made me think, I have no clue how this works – sometimes world class results, while the next minute it seemed like a complete loose cannon. Some good research here and there. No real leverage though.
But sitting at my desk, the project pressure peaking, I figured we might as well see what it could do with real heat transfer calculations. This was the paid version of ChatGPT, but nothing special beyond that. Not customized. Not trained on anything of ours. Just the plain tool.
So I set up the least glamorous workflow you can imagine. Chat window on the screen. Heat transfer textbook open next to it. And we went through the problem together, equation by equation.
It made errors left and right. I mean that literally. It would pull the right equation and botch the math. It would set up a calculation and drop a term. Every time, I caught it against the textbook formulas and my math, corrected it, and sent it back through. It felt exactly like checking a junior engineer’s calc package, except this junior engineer never got tired and never got overwhelmed.
And over those conversations, something happened that surprised me. It started to run clean. It was genuinely good at extracting the equations I needed, good at laying out approaches I could then verify myself. It could track down references, websites, and information that genuinely saved me hours of searching and synthesizing. We worked through the shoe interface, the conduction path, and the real question of whether an insulation barrier was even necessary. I got to a workable, defensible answer inside my two weeks, and I did it without the $30,000 study.
What did the same project teach me about accelerating research and requirements gathering?
This particular project had a second “aha” moment that stuck with me. The facility sat inside city limits, and I needed to know whether I even had to worry about seismic loads for the stress analysis at that geographical location.
The old version of me spends an afternoon in Google and the online code books. Instead, this time I ran it through Perplexity. It searched exhaustively and found the local codes in enforcement, the active publicly available published codes, the risk categories, and a synthesis of the findings and recommendation, all with citations I could click and verify for each step of the process. Just like a junior engineer.
But it also told me something I wasn’t looking for. This city was adopting a new building code edition, flipping over that July. When a new code edition gets adopted, that’s a big deal for engineering, consulting, and construction. It changes your design basis. I wasn’t aware of it. I hadn’t asked about it. The tool surfaced it anyway.
That’s the part of AI research nobody talks about. The speed is nice. Catching the thing you didn’t know to ask about is worth even more.
Where does the engineer fit in all of this?
People hear this story and take the wrong lesson from it. They think the point is that AI did a thermal analysis. It didn’t. I did the analysis. AI accelerated the leg work and synthesis of the analysis.
I pulled from my textbook and other trusted sources. Every equation got checked. Every result got compared against calculations I ran myself. When the model was wrong, and it was wrong more than once, I was the one who found it. Take me out of that loop and you don’t have a faster study. You have a pile of confident, wrong numbers that could put someone in danger.
That’s why I tell engineers the expert in the loop isn’t a safety add-on when working with AI. It is a strict requirement for everything we do. Plan the work with the tool. Do the work with the tool. Then check the work like the skeptic your license requires you to be. The true value was I could focus my efforts on the creativity, decision making, and judgement that only an expert human can do while AI expedited the mundane, time-consuming tasks that don’t require a human. A specialist analysis that didn’t seem feasible became two weeks of focused work at my own desk.
After that project, the tools moved from a curiosity to a daily habit. The guy who thought this thing was a newer version of Google now starts almost every piece of work in a chat window.
My desk is still stacked with textbooks and references. Ready for my peer checks and reviews.
If you’re an engineer staring down a problem you’re not sure how to solve, I can’t guarantee AI will be a silver bullet for you situation. I can assure you that if you use it properly, you will be able develop new solutions, in faster time, and with higher quality than previously possible alone. AI is an amplifier of creativity and expertise, affording us the ability to solve otherwise insurmountable challenges.