Wildfire RFM: Using foundation models to predict wildfires (kumo.ai)

At the KumoRFM August hackathon, Qun Wang and team presented WildfireRFM, a three-layer AI system that uses the KumoRFM foundation model to predict wildfire risk, directs focused real-time monitoring, and fuses observations into automated emergency reports. The pipeline begins with a long‑term predictive layer that models 26,000+ California grid points (nested JSONL) as relational entities (GridPoint and Fire_Event), builds an automatic relational graph via the KumoRFM SDK, and issues forward-looking Predictive Query Language (PQL) queries to produce interpretable per-grid risk scores. That narrows statewide surveillance to Top‑K hot zones, addressing the speed and scale limitations of manual monitoring and opaque regressors. In targeted zones, an Edge AI layer runs a fine‑tuned YOLOv8 detector (trained 100 epochs on Roboflow‑annotated fire/smoke images) and a quantized LLaMA3.1 model on NVIDIA Jetson hardware, fusing detections with live meteorological data (OpenWeatherMap) for instant alerts. A cloud Decision Intelligence layer (gpt-4o-mini) performs multi-source fusion and auto-generates HTML/JSON Fire Risk Assessment Reports with color-coded risk, reasoning, and action recommendations. Experiments showed robust YOLOv8 precision/recall for real‑time detection and balanced LLM risk scoring with fewer false positives in urban settings, demonstrating a deployable, interpretable feedback loop that can speed response, concentrate resources, and improve wildfire situational awareness.