Sentinel Ember

TRY IT LIVE AT: sentinel-ember.vercel.app Sentinel Ember: live wildfire intelligence from AI-piloted drones for quicker wildfire suppression. In the Palisades Fire, the first public fire map came 26 hours after it started. Sentinel Ember cuts that wait to under a second by launching a Muse Spark agent swarm to fly drones over the fire, telling firefighters two things: where the fire is right now, and where it will be in an hour. This way, crews can get ahead of fires before it spreads rapidly. Our fine-tuned YOLOv8 detector runs at the edge, in the browser on the ground-station laptop (ONNX + WebAssembly), and boxes flames in under a fifth of a second per frame. Meta's SAM 3.1 then traces each fire's exact outline. The fire becomes a 3D digital twin on the real hills, with the real wind. On top runs a multi-agent Muse Spark swarm: a forecaster predicts where the front will be in an hour, and a commander and three squad agents assign six simulated drones waypoints to monitor that edge. We test those decisions in a Gymnasium + MuJoCo environment built from real terrain, with thrust, wind drag, battery, camera coverage and line of sight. Built by Jonathan, Dhilen and Ben (Computer Science @ UC Berkeley). Prior work: the base fire detector model came from an earlier project of ours. Everything else was built during the hackathon.