Shape Foundation Model (masked-token pretraining on CAD meshes)
Masked-token pretraining on CAD meshes achieves 0.729 R² reconstruction.

Beats 7B DINOv3 on depth estimation with 1B params using boundary-centric self-supervision.
ML researchers, computer vision engineers, robotics developers
DINOv3 · Segment Anything (SAM) · Depth Anything
Masked-token pretraining on CAD meshes achieves 0.729 R² reconstruction.
Single-binary Rust DB fusing HNSW and BM25 without cloud dependencies or API keys.
Peer-to-peer agent coordination with human checkpoints, but execution unclear and zero traction.
Per-query α fusion beats fixed hybrid weights on FiQA and FEVER benchmarks.
Assembly-specific physics simulation for VLA training when Isaac Sim exists for general robotics.
Surface system auto-bumps depth levels for nested panels without manual z-index management.