The problem
Quantity surveying is slow because it is manual. Someone opens a drawing set, counts what is in it, and writes the numbers down. On a large job that is days of work, and two surveyors rarely arrive at the same figure.
What I built
An agentic system that reads the drawings and documents and produces the take-off, with every number traceable back to the page it came from. The surveyor checks the output instead of starting from nothing.
I owned the AI side of this product end to end, from the first design through to production, between May 2025 and January 2026.
Stack
Python, LLM agents with tool use, document parsing, FastAPI.