LatticeLabs Toolkit

The LatticeLabs Toolkit is a monorepo of seven Python packages that bring 3D CAD geometry into the AI ecosystem — parsing CAD files, tokenizing geometry, running neural models over STEP/B-Rep data, segmenting B-Rep faces, recognizing geometry for LLMs, processing point clouds, and generating parametric CAD. The neural models train and run natively in MLX on Apple Silicon as well as PyTorch.

Each package is independently installable and documented here. The pages below describe what actually exists in the code today — maturity is called out per package so you always know what is production-ready, what is proof-of-life, and what is planned.

Packages

cadling CAD document processing (docling-inspired): STEP/STL/BRep/IGES parsing, topology analysis, RAG chunking, and synthetic Q&A generation.

ll_stepnet Neural networks for STEP/B-Rep files: tokenizer, feature extractor, topology builder, transformer+GNN encoder, and task heads. Trained classifier (0.976), native MLX.

ll_brepnet B-Rep face segmentation (UV-Net + coedge message passing). Trained on the full Fusion 360 split — test mIoU 0.828. Native MLX (parity-verified).

geotoken Geometric tokenizer with adaptive quantization — mesh, command-sequence, and B-Rep graph tokenization for transformer models.

ll_ocadr Optical CAD Recognition: encode global + tiled-local geometry into an LLM's embedding space. HF-native inference today; vLLM serving is experimental.

ll_gen Generation orchestration: neural propose, deterministic dispose in a kernel sandbox. DeepCAD-trained program-based generators produce measured-valid CAD (AR 0.914 / latent diffusion 0.934).

ll_clouds Point-cloud processing (NumPy/SciPy): PLY/PCD/XYZ I/O, preprocessing, normals & curvature, ICP registration, RANSAC/Euclidean segmentation.

How the packages fit together

A typical flow

A CAD file enters through cadling (parse → CADlingDocument). From there geometry can be tokenized by geotoken, encoded by ll_stepnet, fed to an LLM via ll_ocadr, sampled as a point cloud by ll_clouds, or used as a target for ll_gen to generate new parametric CAD against.

Project status

Recently shipped

ll_brepnet — the B-Rep face-segmentation network — is built, trained, and documented: test mIoU 0.828 on the full Fusion 360 split, with a parity-verified native-MLX port. See its docs and the Roadmap. Trained

ll_gen generation now produces measured-valid CAD via the construction-program route (autoregressive command generator + latent diffusion), trained on the DeepCAD distribution and gated on real non-degenerate solids. Generates valid CAD

Still experimental: vLLM serving for ll_ocadr (the HF-native path works today).

Planned