Fire AI
A computer-vision early-warning system that spots smoke in wildland camera feeds minutes before a human operator would.
Problem
A regional fire authority had 140 cameras and three people watching them. Smoke plumes were being caught late, and the difference between a six-minute and a sixty-minute response is the difference between an incident and a disaster. Existing commercial systems assumed new hardware the authority could not fund.
Solution
We built the detection model to run against the feeds they already had, tolerating low frame rates, weather artefacts and poor night contrast. The operator dashboard was the harder half of the problem: alerts had to be glanceable, dismissible in one action, and never so noisy that operators started ignoring them.
Detections stream to the dashboard over WebSockets with the triggering frame attached, so an operator can confirm or dismiss in a couple of seconds without leaving the screen.
Result
The system detects a median of six minutes ahead of manual reporting while holding false positives to under two per camera per week — low enough that operators trust the alerts. It runs entirely on the authority's existing camera estate.
Stack
- Next.js 15
- Python
- PyTorch
- WebSockets
- Cloudflare Stream
- Postgres