See the person in the cell — and why the cell is running slow.
A camera positioned to see whether someone is inside the fence line already sees how long the cycle took and where the queue built. This layer reads both from the same feed — as a monitoring layer beside rated safety control, never in place of it.
What it catches
- A person inside the cell boundary while the cycle is armed
- An obstruction left in the robot's path between cycles
- Gates, guards or light curtains defeated or propped open
- Missing PPE at a station where site procedure requires it
- Cycle time drifting against takt, and which station is causing it
- Starvation and queueing at a station before it becomes a stoppage
Where it reports
- Operator HMI and andon displays as a reviewable event
- EHS incident records with the triggering evidence
- Supervisor alerting through existing channels
- MES and shift reporting for cycle time and station state
- Bounded telemetry to the PLC through a Zephyr companion
- A retained record showing what was seen and what was proposed
What actually runs on the device
- Task
- VLA against a written operating rule, plus classical CV for timing and counts
- Typical target
- NVIDIA Jetson or QCS6490 reference design
- Mode
- Continuous, with event-driven record creation
- Reach
- Whatever you grant on that cell. Read-only by default
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
Is this a replacement for light curtains or safety relays?
No, and it must not be treated as one. Rated safety devices carry a certified performance level; a vision model does not. Indagate is a monitoring and evidence layer that runs alongside those devices. Even where you grant an agent a write path, it is not a safety function — safety relays, interlocks and commissioned procedures stay authoritative.
Why put safety and throughput on the same camera?
Because it is the same view. A camera placed to see whether a person is inside the fence line also sees how long the cycle took, when the station starved and where the queue built. Reading both from one feed means one install, one admission record and one set of evidence — and it lets a safety deployment carry a productivity case without a second project.
Does it perform facial recognition?
No. The models detect the presence and position of a person, not their identity. No face recognition or identity inference is performed, and the decision record contains no biometric data.
How does it avoid alarming on every normal entry?
The operating rule is written per cell and includes the states that make an entry expected — maintenance mode, cycle stopped, a specific gate open. Events are only raised when the observation and the declared state disagree.
What is the latency?
It depends on the device and model you admit. A compact model on a Jetson-class target typically returns a structured result in well under a second on an event trigger. Because this is a monitoring layer rather than a control loop, the meaningful figure is time-to-alert, not cycle time.
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.