Projects · Built
Urban zoning intelligence
An urban intelligence platform that answers property and zoning questions across several major US cities.
The problem
Zoning rules decide what can be built on a lot, but they live in dense, scattered documents and datasets. Answering a simple question about one parcel can take hours.
The approach
- Fine-tuned language models with parameter-efficient training, grounded by retrieval over millions of property records.
- Hybrid search that mixes keyword and dense retrieval, followed by a cross-encoder reranker.
- Pipelines for property classification, zoning prediction, and spatial feature extraction.
- A GIS and 3D layer for building geometry, setbacks, floor-area ratio, and density.
The result
A question about a lot gets an answer grounded in the rules that apply to it, served in real time from GPU-backed infrastructure.