Advancing the frontier of 3D generation, spatial reasoning, and physically-grounded intelligence.
language-conditioned CAD synthesis
whisker — closed-loop parametric B-Rep generation with visual grounding
We present Whisker, a proprietary multi-agent system for generating verified parametric boundary representations from natural language. Our approach decomposes design intent into geometric primitives, constraint graphs, and interdependent parameter spaces, producing STEP-compliant B-Rep solids. A closed-loop verification stage scores geometric fidelity across multiple projections against the original specification, iterating until convergence. We further introduce a proprietary inverse-CAD module that recovers editable parametric programs from unstructured triangle meshes. A novel intermediate representation enables lossless bidirectional transfer between Whisker and the designer's CAD editor of choice — models round-trip without information loss, preserving full feature trees, resolved topology, and linked parameters in both directions.
2026
spatial kittens
multi-agent orchestration · visual grounding · inverse CAD · B-Rep synthesis
2D-to-3D scene reconstruction
houseview — lifting architectural plans into navigable 3D environments
HouseView addresses the ill-posed problem of recovering dense 3D structure from a single 2D architectural floor plan. A proprietary perception stage performs semantic room segmentation and spatial topology extraction, which conditions a generative 3D process to produce watertight meshes with coherent room geometry. A parallel branch synthesises high-resolution equirectangular panoramic projections per detected room, producing seamless 360° environments suitable for stereoscopic rendering. Proprietary camera trajectory planning enables procedural flythrough generation with physically plausible motion interpolation across the reconstructed scene.
2026
spatial kittens
scene reconstruction · equirectangular synthesis · 3D diffusion · trajectory planning
spatial layout optimisation
enso — constraint-driven spatial layout via multi-agent reasoning
We introduce Enso, a system that jointly solves design preference elicitation, product retrieval, and constrained 3D layout optimisation. A conversational agent extracts a canonical design brief through natural interaction, which conditions a proprietary semantic retrieval module operating over real-world product catalogues. Spatial placement is formulated as a constrained optimisation problem: decision variables encode position, orientation, and scaling for each asset; hard constraints enforce non-overlap, boundary containment, and accessibility clearances; the objective maximises a composite aesthetic-functional score. A vision-grounded validation step verifies the solution against the floor plan prior to rendering an interactive 3D scene.
2026
spatial kittens
constrained optimisation · semantic retrieval · multi-agent systems · layout generation