Logarithms
Studio
Specialist models are only as good as the products that carry them. This is the technical shape of each.
DecorLM
Chat-native interior design as a durable workflow. Rooms are reconstructed from LiDAR, photos, or panoramas — VGGT + SAM-3 point clouds into SpatialLM-family floor-plan extraction, the interface our research improves — then layout archetypes are scored against circulation and openings, 2D layouts are generated inside an eval loop with hard metrics, and sourcing constrains every item to real, purchasable furniture with live pricing. Read the essay.
Scaffold
A natural-language 3D engine in the browser: a Three.js/WebGPU viewport backed by cloud-GPU generation — text or image to PBR mesh in three quality tiers (~10s to ~90s) with quad remeshing — Neo4j scene graphs for spatial queries (“what fits within 2m of the fireplace?”), and scene reconstruction that runs analysis, parallel object generation, engine-agent assembly, and a VLM render-compare evaluation loop.
Cobots
One agent per job, onboarded to your tools with least-privilege access, executing skills that are tested workflows with per-step evals. It returns work in your house style, declines jobs outside its competence, and routes them to an agent that can do them. Why we think this is what the interface for AI should be: read The Interface Question.