Inside Vatar’s detection models: how we cut false alarms by 80%
A look at the confidence scoring and scene context that let Vatar tell a real threat from a passing shadow - and why fewer alerts means faster response.
The hardest problem in security AI is not spotting a person, a vehicle or a weapon - modern models do that well. The hard problem is knowing which of the thousands of things a camera sees each day is actually worth waking someone up for. Get that wrong and operators drown in noise; the one alert that matters is lost among a hundred that don’t.
Over the last two quarters we rebuilt how Vatar scores an event before it ever becomes an alert. Instead of reacting to a single frame, the model now weighs scene context - time of day, the zone a subject is in, direction of movement, and whether the behaviour matches a pattern you’ve told us to watch for. A figure crossing a fenced perimeter at 2am is treated very differently from the same figure walking through a gate at noon.
The result, measured across live deployments, is roughly an 80% drop in false alarms with no loss in true-positive detection. Fewer false alarms is not just a comfort metric: every alert that doesn’t need a human frees the team to act on the one that does, and verified threats now reach a responder faster than before.
This is the same philosophy behind everything we build - turn the cameras you already own into security that prevents, not just records. We’ll keep sharing what we learn as the models improve.