— Wildfire detection · Edge AI · Real Time
Autonomous wildfire detection.
Embedded. Real-time.
No cloud dependency.
NVIDIA Jetson Orin NX · C++ / DeepStream · L1 → L2 → L3 Kill Chain
Any camera, any RTSP stream. A single board. An AI system that watches the horizon and decides — autonomously — whether smoke on the skyline is a wildfire starting or just another cloud. Designed to scale to multiple simultaneous streams.
NumidAI detects weak signals that a human would miss, then applies a multi-stage verification pipeline to separate real threats from false positives — rejecting an industrial chimney that looks identical to wildfire smoke.
This is a demonstrator. Terrain validation is ongoing.
Demo
Two real-world cases: prescribed burn detection · industrial plume rejection
Architecture
L1 — Sentinelle
Wide scan · Low threshold
Full-frame detection at 30 FPS. Calibrated to detect weak signals — 12-pixel smoke plumes at 15 km. Everything suspicious is tracked. Most detections are noise. That is by design. L1 is a net, not a filter.
L2 — Sniper
HD crop · Multi-scale re-inference
Persistent candidates trigger a high-resolution crop. The same model re-analyzes the target at 4× zoom. A real smoke plume holds its score. An artifact collapses to near zero.
L3 — Semantic
Vision Language Model · Source classification
Ambiguous survivors are classified by an on-device VLM. Known false-positive sources — chimneys, industrial exhaust, fog — are named and rejected. What cannot be classified as a known non-threat is treated as a credible signal.