Storygold Research notes
What we learn while training a foundation model for parametric CAD: benchmarks, results, and the occasional release.
GPT-5 Astra makes CAD models that look right. Do they work?
Astra reconstructs parametric FreeCAD models from STEP and STL files with startling fidelity. Then we tried to make the parts move.
GPT-5 Astra makes CAD models that look right. Do they work?
Astra reconstructs parametric FreeCAD models from STEP and STL files with startling fidelity. Then we tried to make the parts move.
Learn the solid, not the file
Some AI models for CAD can't tell that two identical parts are the same. A round trip through Rhino drops a state-of-the-art model from 97% to 4% accuracy without changing the shape.
Self-supervised learning for parametric CAD
AI transformed how we write code. Why hasn't it transformed physical engineering? Presenting Masked Topology Modeling: pretraining on structure the geometry kernel already labels for free.
LoRA fine-tuning for vision-language models
LoRA is touted as being as effective as full fine-tuning. On a parametric CAD dataset, it isn't even close.
How good is GPT-5 at 3D?
Syntactically perfect code, dramatically short on geometry. An empirical analysis of frontier-model performance on parametric CAD generation.
An open-source foundation model for parametric CAD
Introducing BELLA-v1.0-13B: weights, evals, and what it can and can't build yet.
Introducing StoryGold
What we're building, why parametric CAD is the right substrate, and why now.
Get the next one in your inbox.
New research notes and early model access. No noise, no schedule.