OmniSight
Watching 100 camera feeds doesn't scale, so we let operators describe what to watch for in plain English.
The problem
Security teams can't watch dozens of live feeds at once, and motion alerts fire on every shadow, leaf, and headlight. Our client needed to surface only the events that matter, without training a custom model for every scenario.
Our approach
We built a natural-language rule engine on top of in-browser vision. An operator writes a rule like “person in the loading bay after 6pm.” The system checks each frame against it with face and object detection plus GPT-4o vision, and only raises an alert on a real match. Running detection in the browser kept it fast and private, with no per-camera infrastructure.
What we shipped
A working monitoring surface that runs detection in the browser, captures matching frames into a review gallery, and delivers alerts in seconds.
- Plain-English rule matching
- In-browser face & object detection
- GPT-4o vision verification
- Live alerts + review gallery
What we took away
A surprising amount of machine-learning work runs in the browser before you need a server. The natural next step is many cameras with a real backend, retention rules, and an audit trail.
See OmniSight in action
This is a real, working build. Open it and try it yourself.
Visit the live demo →Let's build your version.
Tell us the problem. We'll tell you how we'd ship it.