Fireshift AI
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

FireShift AI is a wildfire decision-support tool for incident commanders. It simulates how a fire spreads from fuel, terrain and wind data, then turns that into a timed response plan: when to evacuate, which routes to open and where to send crews. In the replay of the 2018 Camp Fire, it flags the Paradise evacuation 94 minutes before the fire front arrives. It was validated against the real recorded perimeter, with 0.637 IoU and a 96.5% acreage match. A 6-run wind ensemble shows where the fire goes in every run and where it only might, so commanders see uncertainty instead of false confidence. Meta’s models drive the pipeline. SAM 3.1 segments the fire from satellite imagery. Muse Wonder and Muse Spark power an agentic chat that re-runs the simulation live for “what if” scenarios, such as a wind shift. The system also drafts WEA-style public alerts, built to plug into IPAWS and reverse-911, and an Anywhere mode can simulate a fire at any US location.