Architecture

01Decide locally, learn centrally

The camera-to-PLC path runs entirely on the edge PC with a frozen model bundle. Nothing on that path depends on a network, a cloud service or a person. Training, labelling, monitoring and release management live on the central platform.

02The vision system decides; the PLC acts

Reject and divert are executed by the PLC on a hard-wired signal with a watchdog. The AI never drives an actuator directly, and the watchdog acts on the absence of a heartbeat rather than on anything the station has to remember to send.

03No silent PASS

A part is passed automatically only when the evidence is valid — correct recipe, usable views, healthy models, on-time result — and the frozen policy says PASS. Every other case is a HOLD, and a person sees it.

04Everything that affects a decision is versioned together

Model, pre-processing, class map, thresholds, policy and camera recipe ship as one signed bundle with one hash. Each station records which bundle made each decision, so a decision can be reproduced rather than merely remembered.

05Evidence before actuation is complete

The image, the scores, the action and the PLC acknowledgment are written to a durable local buffer before the transaction is considered complete. Upload is asynchronous, and a station that cannot upload keeps inspecting.

06Secure by zone

IEC 62443-style zoning. The cell initiates every connection outbound; there are no inbound ports, no plant credentials on the edge, and only signed artefacts are ever loaded.