Behavior detection
It knows a reach from a pocket.
Vaping on shift. Items slipped into a pocket. Customers who walk out without paying. Each moment is flagged, checked twice, and sent to you with the clip.
“Person raises a small device to the mouth; visible exhale.”
Three behaviors. One careful process.
Vaping
Vaping on shift.
A hand going to the face again and again in a short window. Behind the truck, by the back door, or anywhere you choose.
Concealment
Item slipped into a pocket.
A hand moving to a pocket or the hip and back, in the areas you draw. Built for counters and shelves.
Walkout
Walked out without paying.
Waited at the counter, then left by the door. If Square shows a sale in that window, it never reaches you.
From moment to phone
Four steps. Two of them are checks.
- Our cloud
A moment is flagged
Body-pose tracking spots the pattern on the camera's stream and cuts a short clip around it.
- Video model
The clip is scored
A video model trained on clips owners confirmed or rejected. Clear misses stop here.
- Second check
An AI reviews the frames
Yes, no or unsure, with a one-line reason. It describes actions, never who someone is.
- Your phone
You get the clip
Yes and unsure alert you. No lands in “Filtered by AI”, still there to review.
Gets better every week
Your tap is the training.
Every “correct” or “not correct” becomes a label. Each week a new model is trained on them. It runs quietly next to the current one, and goes live only after we've checked its numbers.
Staging sessions
Teach it on purpose.
Your team acts a behavior out on camera for 15 to 60 minutes, with everyone's consent. Nobody is alerted, and every clip becomes a labeled example.
Leads, not verdicts
Nothing is proof until a person has looked.
Never identifies customers
It describes actions. No customer face data is stored.
Fails open
If a check is down, you still get the alert, marked “not verified”.
Your data, your call
Opt out of training at any time.