Whenever I talk about AI in manufacturing, someone in the room eventually leans back, folds their arms and asks: "Fine. But what did you actually use?"
I've always liked that question. It's the engineer's version of "show me the plant." Slides about "AI transformation" are easy to make. A list of tools that survived contact with real drawings, real deadlines and a real customer's laptop is harder to fake.
So before I share the lessons from the last four weeks, here is the machine itself.
What RevE is, in one breath
SandurTech RevE reads fabrication drawings for heat exchangers and pressure vessels. These are dense A1 sheets crowded with views, sections, a parts list of a hundred-odd rows, design data, a nozzle schedule, notes and a title block. RevE turns all of that into structured, checked data that an engineer can review in an afternoon instead of re-typing over two days. It then does the part my friend at the fabrication unit cared about most: it makes the next drawing, using an existing one as a template.
Version 1.0.1 went out in October 2026, and the Windows build now comes straight out of GitHub at the press of a button. It's a desktop app, because that is how a fabrication office works. One shared laptop, several people, drawings arriving by email.
The surprising shape of it
If you drew RevE as a block diagram, the AI would be the smallest part of it.
Of roughly thirty steps between "open the PDF" and "here is your checked catalogue", only three kinds of work go to an AI model: finding the regions on a sheet, reading text that exists only as drawn curves, and judging engineering meaning. Everything else is ordinary code, the kind you can test, debug and trust: geometry, arithmetic, checking, record-keeping.
I'll explain in Part 1 why that ratio is deliberate. For now, here are the tools.
Choice of tools for SandurTech RevE
Five tools that mattered
1. PyMuPDF, the unsung hero. This open-source PDF library did the unglamorous half of the work. It gives us every piece of text with its exact position, every line the draughtsman drew, the page rotation and fast rendering. I went in thinking the hard problems would be AI problems. Most of them turned out to be PDF problems, and this library quietly solved them.
2. Two OpenAI models, not one. We use gpt-6.1-sol (about US$2 per million input tokens and US$10 per million output) where judgment and spatial reasoning matter. We use the much cheaper gpt-6-luna (US$0.10 and US$0.50) where measurement showed it reads just as well. Choosing per task, rather than picking one "best" model, roughly halved our cost per drawing. That story gets its own article.
3. A local OCR engine. RapidOCR runs right on the laptop, with no internet and no cost per call. On drawings where every letter is a drawn curve, it finds “where” the text sits, so the AI spends its effort on reading rather than searching.
4. ezdxf, the bridge to AutoCAD. Fabricators live in AutoCAD, so new drawings go out as DXF files. When we draw a DXF back and compare, more than 95 percent of the original ink is there. That number mattered more to the shop than any AI accuracy figure.
5. A light desktop shell. pywebview gives us a native window, and three.js draws the exchanger in 3D at true scale. To the user it feels like a familiar desktop application. It runs comfortably on a 4–8 GB Windows laptop.
Two quieter helpers joined later. An open-source spell checker (pyspellchecker) now finds typos printed on old drawings, alongside a list of engineering words so that BAFFLE and TUBESHEET are never "corrected". And GitHub Actions builds the Windows version of the app in about three and a half minutes, so I can develop on a Mac and still ship to the laptop the customer actually owns.
Four things we chose not to use
This is the part I'm proudest of, because every one of these was tempting.
A local AI model. It's the standard answer to "we can't send our drawings to the cloud." I wanted it to work. But a vision model good enough to read engineering drawings simply doesn't run on the laptop this customer owns, and pretending otherwise would have meant a demo on my Mac and disappointment on theirs. We protected the data in other ways, which I'll cover later in the series.
An agent framework. Agents are the fashion of the year. But when we listed what each "agent" would actually do, every step was a plain function with clear inputs and outputs. A framework would have added moving parts that a small fabrication shop can't maintain once I hand it over.
A vector database. Everything RevE extracts from one drawing fits in a few hundred kilobytes. Simple keyword search on the laptop answers most chat questions ("shell thickness?", "list the nozzles") in about a millisecond, for free, without anything leaving the machine.
Paid CAD toolkits. My friend works in AutoCAD and would have liked native DWG files. The libraries that write DWG properly come with licence fees that a small shop's product can't carry. We tested the free options, weren't happy, and made a firm decision: RevE writes clean DXF, which AutoCAD opens and edits perfectly well, and uses free software only. A product priced for small fabricators can't be built on components priced for large ones.
Saying no to fashionable tools is not conservatism. It is respect for whoever has to live with the system after the consultant leaves.
The numbers, for the sceptics
These numbers come from testing the application with real sample drawings and can vary with the drawing sizes and complexities, provided here for indicative purpose only.
- US$0.32 in model costs to read a complete drawing, down from US$0.69 in earlier versions
- US$3.84, the total cost of the experiment that decided which model does which job
- Zero AI calls to generate a new drawing from a template
- 1,815 of 1,815 parts-list fields correct across five real drawings, after review
The real secret isn't in the list
If you copied my five tools tomorrow, you wouldn't get RevE. What made it reliable was the way of working.
Every change was a small, numbered task with its own automated check. Nothing counted as done until it had run on real drawings, on the real machine, and the result had been looked at by a person. And one rule held above all others: if the drawing doesn't contain a value, RevE says so. It never fills the gap with a guess.
I spent decades in steel plant automation, and that rule would have saved more than one commissioning weekend. In engineering, an honest blank beats a plausible wrong number every time.
Next, in Part 1: three of our five drawings had no text at all. How do you read a drawing like that?
#IndustrialAI #Manufacturing #EngineeringDrawings #SandurTech #AIArchitecture #Fabrication #HeatExchangers #OpenAI #Python #DigitalTransformation
Disclaimer
This series introduces SandurTech RevE, a product in development for AI-assisted engineering drawing analysis and generation. Examples reflect our current capabilities and ongoing validation. Outputs should be reviewed by qualified engineers and verified against approved drawings, standards, and project requirements before use.