Hierarchical Reconstruction Models for Industrial CAD

We do not use traditional LLMs because industrial CAD design does not allow for uncertainty or 'hallucinations': it requires geometric precision and deterministic mathematics. That's why we are developing a Hierarchical Reconstruction Model (HRM), an architecture designed specifically to transform implicit AI geometry (meshes, voxels, SDFs) into parametric, editable CAD ready for manufacturing. The HRM operates through hierarchical levels: first identifying fundamental geometric primitives, then topological relationships and constraints, and finally reconstructing explicit parameters (radii, tolerances, symmetries, assemblies). This approach eliminates probabilistic ambiguity and enables reliable conversion from generative models into real engineering. This layer is the critical bridge between generative AI and industrial manufacturing, currently non-existent in the market. Furthermore, the mathematical core of the HRM —based on recurrent, local, and structured operations— is exceptionally compatible with photonic accelerators, unlike dense LLMs. This enables nearly instantaneous geometric reconstruction with minimal energy consumption, positioning the system for the next generation of hardware.