An industrial safety deployment
A computer vision safety system designed to spot smoke before a conventional sensor can. No new hardware, and no video leaves the site.
Conventional smoke detectors need smoke to reach the sensor. In a large open space, a warehouse, a plant floor, a yard, that can take minutes. In a high-ceiling or well-ventilated environment it can fail entirely. The site already had extensive CCTV coverage. The question we are answering is whether those cameras can see smoke before a sensor could smell it.
A computer-vision smoke detection system that runs against existing camera infrastructure: no new hardware, no re-cabling.
CCTV feed
Existing cameras · no re-cabling
Model watches motion
How smoke moves, not what it looks like in one frame
Tuned per camera
Each scene is different · one model doesn't fit all
Alert with the clip
Lands in the existing escalation path · verify in seconds
Operator review loop
Confirm or dismiss · feeds back into training
Detection that recognises smoke by how it behaves over time rather than how it looks in a single frame — because smoke is defined by motion, and a still-frame model confuses it with steam, dust and light shafts. Each camera is calibrated separately so a model tuned to one scene doesn't miss events or cry wolf in another.
The whole model runs on-site, so video never leaves the premises and detection doesn't depend on an internet link — a safety system that fails when the connection drops is not a safety system. Alerts land in the site's existing escalation path with the triggering clip attached. An operator review loop feeds back into training. Camera health monitored across the estate, because a safety system silently watching a dead stream is the worst possible failure mode.
Active development. Architecture decided, detection model in training, edge pipeline running against sample feeds. Pre-production: not yet deployed to a live site at operational scale.
Early visual detection across the camera estate, running at the edge, with every alert carrying the clip that triggered it. Measurable detection latency ahead of conventional sensors in environments where sensors are slow or unreliable.
Capabilities used
Tell us about it. We'll tell you honestly whether we're the right team to help ship it.