Does this use its own backend/engine? I've been working on LLM to CAD tool[0] and have realised there are so many backends and options to choose from. Since the realisation I'm trying to find the best representation for an LLM. I think OpenSCAD is currently the best and most feature complete choice, but I definitely need to dig a bit deeper. If anyone has any pointers I welcome them!
> I think OpenSCAD is currently the best and most feature complete choice
As much as I love OpenSCAD, I would strongly disagree with your conclusion.
All the OpenSCAD language can do is boolean operations and moreover, the engine can only implement those on polygonal (triangle actually) meshes.
That's a very far cry from what a modern commercial CAD engine can do.
For example, the following things are very, very hard to do, or even specify using OpenScad:
- Smooth surfaces, especially spline-based
- Fillets / Chamfers between two arbitrary surfaces
- Trimming surfaces
- Querying partly built models and using the outcome in the subsequent construction (e.g. find the shortest segment between two smooth surfaces, building a cylinder around it and filleting it with the two surfaces, this is an effing nightmare to do within the confines of OpenSCAD)
- Last but not least: there is no native constraint solver in OpenSCAD, neither in the language nor in the engine (unlike - say - SolveSpace)
I might have misunderstood what you're looking to do, but, yeah, digging deeper feels very much like the right thing to do.
using BOSL2 alleviates most issues I've run into with OpenScad for chamfers and the like, but it is an extra set of functions you need to remember sadly
> BOSL2 ... but it is an extra set of functions you need to remember sadly
It's also extremely slow: it implements chamfers and fillets using morpho, and if you have a large number of fillets, the morpho algorithms (minkowski / hull) are very much non linear in time on polygonal meshes, which leads to compute time explosion if you want a visually smooth result.
This "screenshot -> refine loop" is a great strategy and I have built it into my 3D Modeling product as well[0], but had to disable it because it would often quadruple the costs and the product is already expensive.
I am on standby to enable it though, just need a price to drop a bit more!
My late maternal grandfather was Slovenian, so I enjoyed your project's backstory. I've mucked around with ChatGPT and OpenSCAD so can identify with that also. Great concept and best of luck!
In the coworking I am in people are hitting limits on 60$ plan all the time. They are thinking about which models to use to be efficient, context to include etc…
I’m on claude code $100 plan and never worry about any of that stuff and I think I am using it much more than they use cursor.
Tell them to use the Composer 1.5 model. It's really good, better than Sonnet, and has much higher usage limits. I use it for almost all of my daily work, don't have to worry about hitting the limit of my 60$ plan, and only occasionally switch to Opus 4.6 for planning a particularly complex task.
Ran a bunch of 3D Modeling benchmarks on Gemini 3.1 vs Gemini 3.
Unsurprisingly 3.1 performs a bit better. But surprisingly it costs 2.6x as much ($0.14 vs. $0.37 per 3D Model Generation) and is 2.5x slower (1m 24s vs. 3m 28s).
To me it feels like "lets increase our thinking budget and call it an improved model!"
I am building pretty much the same product as OP, and have a pretty good harness to test LLMs. In fact I have run a tons of tests already. It’s currently aimed for my own internal tests, but making something that is easier to digest should be a breeze. If you are curious: https://grandpacad.com/evals
Dimensionally accurate AI 3D modelling. My grandpa has a 3D printer but struggles to use any complex tools. So I am working on this chat interface to allow him to do some simple models.
So far he has triggered more than 150 generations. It’s getting better every model cycle and gives me something I enjoy working on.
Serial usecases ("fix this syntax errors") will go on Cerebras and get 10x faster.
Deep usecases ("solve Riemann hypothesis") will become massively parallel and go on slower inference compute.
Teams will stitch both together because some workflows go through stages of requiring deep parallel compute ("scan my codebase for bugs and propose fixes") followed by serial compute ("dedupe and apply the 3 fixes, resolve merge conflict").
I've been using 5.1-codex-max with low reasoning (in Cursor fwiw) recently and it feels like a nice speed while still being effective. Might be worth a shot.
Very interesting thanks! I wonder what would happen if you kept running Gemini in a loop for a while. Considering how much faster it ended it seems like there is a lot more potential.
[0]: https://GrandpaCAD.com
reply