FA©26
18 June 2026·3 min readCustom AICADSolidMake

Text to STEP: why SolidMake couldn't be a GPT wrapper

Every few weeks somebody tells me text-to-CAD is a solved problem: point a good model at the prompt, ask for a part, done. I built SolidMake for mechanical engineers, and the first thing the build taught me is why that answer collapses the moment a fabricator gets involved.

Prose is not geometry

Ask a general model for “a 4 mm mounting bracket with two 6 mm holes on a 40 mm centre” and you get an excellent paragraph about a bracket. What a CAD kernel needs is something else entirely: an ordered sequence of construction operations — sketch a profile, constrain it, extrude it, cut the holes — where every step is valid given the step before it.

That sequence is the actual output format. Getting a model to emit it reliably is a modeling problem, not a prompting problem. This is why SolidMake is built on DeepCAD and Text2CAD rather than on an API call — those are models that speak in CAD command sequences, and that is the language the downstream kernel is waiting for.

Normalize the prompt before the model sees it

Engineers don't write prompts. They write the way they'd talk to a machinist: units dropped, tolerances implied, half the geometry assumed from context. Two requests that mean the same part can look nothing alike as strings.

So a normalization layer sits in front of the model — it resolves units, makes implicit dimensions explicit, and pushes the phrasing toward the shape the training data actually uses. It is unglamorous work, and it moved output quality more than any other single change.

  • Units to one canonical system before anything else touches the text.

  • Implicit features (through-holes, symmetry, fillets described in passing) named explicitly.

  • Vocabulary mapped onto the training distribution instead of hoping the model generalizes.

STEP is the finish line, not the export button

The output that matters is a STEP file, because that is the format the rest of the world runs on — the thing an engineer drops into their own CAD tool, or hands to a shop for cutting, bending and welding. A result that looks right on screen and fails to export is not a result.

The test was never “does it look like a bracket”. It was “can somebody make this”.

What I'd tell you if you're building domain AI

  1. Find out what your domain's real output format is, and make the model speak it natively. Anything else becomes a translation layer that leaks.

  2. Spend your first week on input normalization. The gains are boring, large and immediate.

  3. Define success as something downstream and physical — a file that opens, a part that fabricates — not as a plausible-looking render.

  4. Pick your base models for the shape of their output, not their leaderboard position.


SolidMake started as my final-year project and became the deepest technical thing in this portfolio. It's also the honest answer to “can't you just use an LLM for that?” — sometimes, yes. For geometry a shop has to cut, not yet.