Turn camera frames into events a person can actually review.
Most industrial camera estates record everything and surface nothing. This layer reads frames against your written rules and produces events with the evidence attached.
What it catches
- Entry to a restricted area outside permitted hours
- Unattended items left in a walkway, bay or entrance
- Vehicles in the wrong zone or blocking an egress route
- Perimeter and fence-line activity after hours
- Doors and gates held open beyond the configured threshold
Where it reports
- Existing VMS and access-control platforms
- Guard-tour and incident systems with the frame attached
- On-call alerting via webhooks, n8n, Zapier or Make
- A retained decision record with model identity and region
- Argo, so a duty operator can ask what happened and why
What actually runs on the device
- Task
- Compact VLM against a written site rule
- Typical target
- QCS6490 reference design or x86_64 gateway
- Mode
- Motion or schedule-triggered snapshot
- Reach
- Whatever you grant. Read-only by default; door control is a separate grant
Every one of these passes the Oddy test against your named device before it is admitted. A candidate that does not fit the device envelope never receives an edge directive.
Common questions
Does this identify individuals?
No. The security models describe what is happening in a frame — a person present, an item left, a vehicle in a zone. They do not perform face recognition or identity inference, and every output requires human review before action.
Can it unlock or lock doors?
Only if you grant it that reach, and by default you have not. Out of the box the agent raises an event and a person or your existing access system decides what follows. Access control is a separate grant, scoped to named doors, and the reach coupon rejects any agent that asks for it without one.
Where do the frames get processed?
In a customer deployment, on your own hardware by default. You may configure a private region or hybrid route instead, and whichever route is active is disclosed in the decision record along with the model and region.
How is this different from the analytics already in our VMS?
Built-in VMS analytics generally detect fixed classes — motion, line crossing, a person. A vision-language model reads the frame against a rule you wrote in plain language, which covers situations nobody configured a detector for. It also records which model produced the judgement, which fixed-function analytics rarely do.
Start with one line.
Send your SOPs and we will audit them free — the agent that fits your procedures, and the hardware it runs on.