ArchionLabs

Claude parses the prompt and extracts room requirements, with a regex fallback if the response fails to parse. Requirements go to a LayoutSolver that uses a packing algorithm to place rooms inside the building boundary. The project and its chat history persist to PostgreSQL, then sync to the dashboard.

The detail that makes it a product rather than a demo: when an existing floor plan is passed with the request, stage one becomes a modification prompt instead. The existing plan JSON goes to Claude, which returns a revised version, bypassing the solver entirely. Generating and editing share one endpoint, and each revision is stored as a new project rather than overwriting the last.

Floor plan JSON runs 50KB to 200KB, and every modification prompt has to carry the whole object. Claude's 128K window is what makes conversational editing viable at all. Vision handles extraction from uploaded image and PDF plans, and in practice it returned structured JSON without Markdown wrappers consistently enough to parse directly. The service switches between model variants through an environment variable: Haiku for fast iteration, Sonnet where layout quality matters more than latency.

With hundreds of elements in a scene, DOM manipulation cannot hold 60fps. The editor runs a requestAnimationFrame loop at 60fps, stores all geometry in world coordinates in metres, and applies a pan and zoom affine transform at render time so the underlying model stays in real units. Eight tools run off a single Zustand store, chosen over Redux for lower overhead on a canvas that rebuilds every frame.

Two decisions mattered more than their size. Architectural symbols follow ISO 128/7519, so a door renders as a leaf line from the hinge plus a quarter circle swing arc, which is what an architect expects rather than an approximation. Those drawing functions live outside the React tree as standalone canvas functions, which made them independently testable. Undo and redo use structuredClone snapshots capped at fifty entries, so shared mutable references never leak between states.

Each backend is an independent FastAPI application with its own database config, Dockerfile, and test suite, communicating only over HTTP and sharing nothing but the Supabase JWT secret. A failure or deployment in one cannot affect another. Auth is exposed as two FastAPI dependencies: a strict one that raises 401 on a missing or invalid token, and a lenient one returning None for routes that tolerate anonymous use.

The engine models pedestrians as intelligent agents with individual objectives, spatial knowledge, and decision-making capability, rather than as particles driven by hand-fitted physics constants. Deep learning models handle the high-dimensional state space required to run many agents at once, computing pathfinding and collision avoidance in real time as agents move through the 3D geometry.

The reason this matters for the product: a deterministic model only produces the behaviour you encoded, and the entire value of the tool is showing an architect something they had not predicted about their own design. Congestion and bottleneck formation have to be able to emerge rather than be scripted.

The tradeoff is real. Training time replaces tuning time, runs are harder to reproduce, and diagnosing a policy that behaves oddly is far slower than adjusting a constant.

Running to completion before responding would leave the client waiting for the length of the run. Instead the start endpoint spawns the engine in a background thread and returns immediately. The client subscribes over SSE through the browser's native EventSource API and receives frames as they compute, each carrying the position and type of every agent. On a done event, the frontend triggers the compliance and analytics fetches and enables PDF export.

This required infrastructure work alongside application code. Nginx sits in front as a reverse proxy and had to be configured to hold SSE connections open rather than buffering them, which is the default and silently breaks streaming.

Pedestrians are generated humanoid meshes updated per frame through React Three Fiber's useFrame. Position smoothing uses frame-rate-independent exponential interpolation so movement stays natural regardless of playback speed. Walking animation is procedural: leg and arm geometry driven by sinusoidal functions on a per-agent cycle counter. That avoids shipping rigged character assets entirely.