Insurance · 2026 · Self-built prototype
FireScore: a parcel-level wildfire risk score that shows its work
A working proof-of-concept that scores wildfire risk for a single parcel from public authoritative data, explains every point of the score, and serves the same answer to a person on a map and to an AI agent over MCP.
- Sector
- Insurance
- Year
- 2026
- Engagement
- Self-built prototype
- Team
- Solo build, with Claude Code
We built FireScore to answer one insurer question honestly: how exposed is this single address, and why. Not a heat map you squint at, but a number a person can defend, decomposed into the fuel, the fire history, the terrain, and the access that produced it. It is a proof-of-concept, built end to end on public data, before we have taken a dollar for it.
What we built
One PostGIS database is the single source of truth. It carries fire perimeters, fuel models from LANDFIRE, hazard-severity zones, slope, and vegetation, all loaded idempotently from public authoritative sources. A calibrated SQL scoring pass turns that into a 0–100 number for any geocoded parcel. Two surfaces read the same data: a MapLibre web app for a human, and a four-tool MCP server so an AI agent can ask for a score and get one back with its inputs attached.
What we found
The score was the easy part. The honest work was provenance: an underwriter will not trust a number they cannot take apart, so every score decomposes into the exact fuel class, perimeter distance, slope, and zone that moved it.
What we learned
The confession here is scale. This is a proof on a hundred parcels chosen to span a hazard gradient, not a carrier's book of a million. The calibration is tuned to a range we picked; a real portfolio will surface distributions we have not seen, and the weights will have to move when it does. We would rather say that now than discover it in front of a client.