The AI Experimentby Xelec

Industrial manufacturing · Germany

Visual inspection on the line, offline, at 38 ms

Cloud inference was impossible on the shop floor. A distilled model on Xelec hardware runs at the station.

38 ms
median inference
0 bytes
leave the plant
-72%
escaped defects

Context

Surface-defect inspection depended on operator judgement at the end of a line running three shifts. Defects escaping to assembly were expensive and inconsistently recorded.

Constraints

Works council and IT security ruled out sending line imagery off-site. Cycle time allowed under 100 ms per part. The station had no room for a workstation-class machine.

What we built

A vision model distilled from 60,000 labelled frames, quantised to run on a Xelec-designed board mounted at the station, with store-and-forward telemetry and OTA fleet updates across eleven lines.

Outcome

Inference settled at 38 ms median. Escaped defects fell sharply against the prior quarter baseline. No imagery leaves the plant network.

Illustrative engagement pattern. Client identity withheld; figures are representative, not audited.

Next step

Start with a conversation, not a contract.

Thirty minutes with an engineer who has shipped this before. If there is no fit, we will say so on the call.