Parametric Massing Generator
Zoning-to-form generation · Three.js, Claude · parcel geometry through urban block scale
The Parametric Massing Generator is a browser-based computational design system that translates zoning language into inspectable building envelopes. It asks whether AI can assist with the first stage of architectural massing without turning regulatory interpretation into an opaque act of trust. Every generated value is paired with a source quotation and confidence level, every form is subject to a separate review pass, and every revision remains visible in a traceable history.
Three Scales of Massing
The tool operates across three distinct design problems: simple lot mode for rectangular parcels with setbacks, height, and floor area limits; irregular parcel mode for non-rectangular sites; and urban block mode for subdivided sites whose parcels are assigned residential, commercial, administrative, or park functions. This progression connects individual building envelopes to larger questions of urban form rather than treating zoning as a single fixed number.
System Architecture
When a user provides a bylaw or plain-language description, Claude returns eight structured constraints: lot width, lot depth, front, side, and rear setbacks, maximum height, floor-to-floor height, and maximum floor area ratio. Each field includes the exact supporting phrase and a confidence level. A second, independent review call checks the original text, extracted values, and generated statistics, then can propose a specific correction. A conversational refinement layer supports instructions such as “make it one storey taller” while changing only the requested field and preserving the rest of the current state.
Geometry and Verification
For convex parcels, the buildable footprint is calculated through inward polygon offsetting by half-plane intersection, allowing the same verified method to handle rectangles, chamfered lots, triangular parcels, and urban-block subdivisions. The system renders a Three.js perspective view beside a north-up, colour-coded plan and records parsing, review, correction, and refinement actions in a generation log. The geometry was tested against hand-computed rectangle and triangle cases, including the expected collapse of a parcel under an oversized setback. A slider-rounding bug that changed an AI-parsed 4.9 metre value to 5 metres was also found and fixed by tightening dimension controls to 0.1 metre increments.
Vancouver R1-1 Case Study
The built-in preset encodes current R1-1 provisions, including a 4.9 metre front setback, 1.2 metre side setbacks, a 10.7 metre rear setback, an 11.5 metre and three-storey height limit, and a base floor space ratio of 0.70. The rental or below-market housing condition changes the applicable ratio to 1.00, while frontage and lot-area thresholds determine the eligible multiplex unit tier. These conditional rules remain explicit instead of being flattened into one representative number.
Project Outcomes
The result is an AI-assisted massing workflow that exposes its reasoning rather than hiding it behind a finished-looking model. The Three.js geometry remains usable without an API key, while AI parsing, review, conversational refinement, and urban-block classification add a layer of assisted interpretation. The project demonstrates a practical division of responsibility: deterministic code computes the form, while AI extracts, critiques, and communicates the evidence behind it.