Inspection, integrated.

Defects escaping a manual line.

Visual inspection on two production lines relied on operators sampling parts at intervals. Sampling meant some defective units escaped to the next stage — expensive to catch later, worse to catch after shipment. A prior vision pilot had produced good accuracy numbers in a lab test but was never connected to the systems that would let it act: no path into the MES, no defined process for what happens when it flags a part.

Built for the plant floor, not the demo.

We started from the integration, not the model. The vision system runs at the edge, inline with existing line-side cameras, and every inspection result is written directly into the MES as a native event — pass, flag, or hold — using the same event schema the plant's other systems already consume. No new dashboard for operators to learn.

Flagged parts route to a human for a fast visual confirmation before a hold is issued, so the system's judgment is checked before it stops a line.

Edge inference, MES-native events.

Inference runs on edge hardware at each inspection point to keep latency inside the line's takt time. Results post to the MES as standard events; a lightweight operator console shows the flagged frame and the confidence for the fast human check. All inspection decisions — model and human — persist to a single audit log per part.

Fewer escapes, same line speed.

Coverage moved from sampled to full inline inspection on both lines without slowing takt time, because inference and the human check both run inside the existing cycle. Escape-defect rate dropped in the first full quarter of operation, with every hold traceable to a specific frame and decision.