Hi HN,
I have a 3D printer but I’m pretty bad at CAD. Using LLMs for coding has worked extremely well for me, so I’ve been trying to apply a similar workflow to CAD/modeling.
For simple functional parts (jigs, brackets, adapters, small fixtures), I can use an LLM + OpenSCAD in a loop: it writes OpenSCAD, I compile/render, I render a few views, the LLM “looks” at the images, and we iterate until it looks right. This is already helpful, but it hits a ceiling quickly. Anything beyond simple parametric primitives becomes painful (complex geometry, precise interfaces, assemblies, tolerances/fit, etc.).
I’m curious about two things. First: any intuition on when we’ll be able to generate models at a professional level, comparable to what LLMs can do for coding right now? Second: what will that workflow look like in practice? Will it stay mostly parametric (OpenSCAD / constraints), or will it look more like an interactive “CAD copilot” inside tools like SolidWorks/Fusion that can edit the feature tree via screenshot + click style interaction? Or something else entirely, like text turning into a full feature history with constraints and checks.
If you’re already doing this, what tools/workflows are giving the best results today, and where do they fail?
Thanks!
I'm working on a parametric, feature based CAD by code software which should be released next month. LLMs are good at generating code. It should be able to handle any complex task given a good documentation. I'm planning to write an mcp for providing feedback to the LLM but that will come at a later stage.
The code is based on js and opencascade with a feature rich vscode extension.
I'll post it here on release day.
Hey there! I'm open sourcing https://campedersen.com/cad0 today! It's not professional-level yet (about a week old) but I think the next model generation could get there
I can barely get AI to generate decent Mermaid diagrams
I've been in the programmatic CAD space for a while, and the core issue isn't geometry generation, it's the feedback loop.
Current approaches hit the same wall: you can't easily go from geometry back to code. When an LLM generates OpenSCAD and the output is wrong, you're staring at an STL with no way to point at a face and say "this edge should be 2mm shorter." The LLM is flying blind.
The tools that work best right now are the ones that keep geometry minimal and code iterative. OpenSCAD with an LLM in a loop can handle simple parts because the code is the source of truth and changes are surgical. But complex assemblies? The boolean operations cascade failures in ways that are brutal to debug even for humans.
The real unlock isn't better prompting or bigger models: it's a representation that supports bidirectional editing. F-Rep/SDFs are interesting here because the math is more compositional and less "undo history dependent" than B-Rep.
Manufacturing constraints (tolerances, printability, material limits) are a separate unsolved problem. Geometry that looks right but won't print is the silent killer.