SARchlight - Multi-Drone Search and Rescue
Terrain-aware multi-drone search with a Bayesian belief map, a live dashboard, and a deployed voice agent. Built with a team at a search-and-rescue hackathon.
Technologies
About This Project
The AI decision layer for wide-area drone search and rescue. A Bayesian, terrain-aware belief grid (built from real elevation and land-cover data) plans disjoint multi-drone coverage, fuses detections with clean sweeps, and declares a subject located only after persistent evidence, then guides them home along a terrain-aware route and speaks to them. A React dashboard reads a FastAPI state API, and voice runs through Deepgram, including a deployed Twilio and Deepgram operator phone agent with a live transcript feed. The full loop runs end to end on a realistic detection simulator, with a real Ultralytics YOLO detector built and unit-tested behind the same interface, ready for footage. It is validated by a large pytest suite spanning unit, integration, and soak tests. Built with a team at the UC Berkeley AI Hackathon in June 2026.