Graph Machine Learing: La Cité Radieuse
Abstract
How does the spatial organization of Unité d'Habitation influence circulation, accessibility, and apartment connectivity, and can machine learning predict room functions from graph properties?
Context / Problem
Industrial layout design is governed by strict standards: clearances, corridor widths, equipment access rules. Yet the process remains manual and validated only after completion, with no system capable of reasoning spatially while the design is still in progress.
Methodology
The team built an agentic pipeline using LangGraph and Claude, grounded in a NetworkX spatial graph encoding rooms, furniture, and relationships as structured knowledge. Six tools (Collision, Reachability, Orientation, Path Analysis, Visibility, Scoring) run in real time, feeding spatial feedback back to the LLM so it can place, adjust, and narrate at each step.
Results
Across 20 sessions, the system achieved a 95% cumulative success rate, a 6.4-minute average run time, and 51 tool calls per workflow. Claude consistently outperformed Gemini in multi-step sequencing, spatial reasoning, and collision resolution.
Conclusions
Spatial Flow demonstrates that a language model given structured spatial knowledge, measurable constraints, and persistent memory can generate and iterate industrial layouts from a single sentence, producing outcomes that are not just plausible, but industry-feasible.
Project Documentation
