Catch the defect on the line, not at the customer.
Four checks that all reduce to the same question: does the part in front of the camera match what the work order says it should be? Device-sized vision models run beside the line and report into the systems you already use. No frames leave the plant unless you decide they should.
What it catches
- Defect detection — scratches, contamination, weld and seam faults
- Assembly verification — missing, misplaced or wrong-orientation components
- Label and date-code inspection — presence, legibility and match to the order
- Packaging integrity — crushed, damaged or incorrectly sealed
- Process anomalies that no one wrote a rule for yet
Where it reports
- PLC tags over a bounded Zephyr companion, scoped to the tags you name
- CMMS work orders with the triggering frame attached
- MES / ERP quality records via governed HTTPS APIs
- Existing alerting through webhooks, n8n, Zapier or Make
- A retained decision record with model identity and evidence
What actually runs on the device
- Task
- Compact VLM, or classical OpenCV where a VLM is overkill
- Typical target
- QCS6490 reference design, NVIDIA Jetson, or x86_64 gateway
- Mode
- Trigger-driven snapshot on part presence
- Reach
- Whatever you grant on that device. Read-only by default, no shell
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
Do we need labelled training images to start?
Not for the first pass. A compact vision-language model can describe what it sees against a written operating rule, which is enough to establish whether the signal exists. Task-specific models are worth training once you know the defect classes matter and how often they occur.
Can it stop the line?
Only if you grant it that reach, and most deployments do not. The default is read-only: the agent produces a recommendation and a record, and stopping the line stays with safety-rated control. Where a write path is granted it is scoped to named tags on a named device, the Oddy test admits the agent against exactly that grant, and every action it takes is recorded. Even then it is not a safety function — interlocks and operators remain authoritative.
What happens to the frames?
In a customer deployment, frames are processed on your hardware and retained according to your policy. Nothing is sent to a cloud provider unless you configure a hybrid or private-region route, and the route in use is reported in the decision record.
Can we use our existing cameras?
Usually. The runtime works with GigE, USB3 and MIPI sources, and with most IP cameras that expose RTSP. Global-shutter cameras with trigger and strobe give better results on fast-moving lines, which is why the EdgeSight C1 exists — but it is not a requirement.
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.