Work

A framework for building AI agents on top of Snaptrude. The first agent lets architects describe what they want in plain English and produces a concept massing in about 10 minutes. Handles zoning research, adjacency generation, and checks its own output. What used to take 4 days now takes about 4 hours.

Real-time collaboration for 3D architectural models. Multiple people can edit the same building at the same time. We stress-tested it with 60 users. No file merging, no sync conflicts.

Open-source tool that lets you share a terminal session in the browser. Create a room, share a 6-character code, everyone can see and type. I built it for sharing Claude Code sessions with my team.

Boolean operations, clipping, and transformations for 2D polygons. Lightweight, pure Python. 38k+ installs on PyPi. Started as internal tooling at Snaptrude and I pulled it out into a standalone library.

When your LLM pipeline crashes at line 25 million, you don't want to re-read everything to resume. This builds a byte-offset index so you can seek to any line in O(1). Python and TypeScript, published on PyPi.

A live spreadsheet interface over 3D scene data. Architects edit cells and the model updates in real time. Turns out they'd rather work with tables than a 3D viewport for bulk edits.

AI rendering that stays true to the actual 3D geometry instead of making up proportions. Uses depth mapping from the model to keep outputs accurate. Built this during the Snaptrude Hackathon 2024.

Architectural models create massive MongoDB documents. I built an indexing layer that narrows the search radius per query and cut query times by 80% across the platform.