Good parts in. PASS, FAIL and a record out.

Seat good parts

Pick one part number. Take good parts from your line, the ones that would ship today, and seat them in the fixture one at a time. The setup wizard takes about 60 photos. You do not need a single bad part to start.

• Same fixture, same light, same angle for every photo
• Good parts only; normal tool marks and finish variation are part of 'good'

Train on the station

The model looks at small patches of every good photo and remembers what normal looks like for this part. Training runs on the station. When it is done, the station knows your part, under your light, in your fixture.

• One model per part number
• Retrain when the part, finish or fixture changes

Inspect every part

Seat a part. The station compares each patch of the new photo with what it learned. Anything that looks unlike every good part raises the score and lights up on the heatmap. Above the threshold the part is FAIL and goes to the HOLD bin.

• PASS or FAIL plus a heatmap of where it looks different
• Flagged parts go to a person, not straight to scrap

A person decides

The inspector looks at the flagged part and the heatmap and makes the call. If they disagree with the station they override it and write a reason. Both verdicts are kept, so you learn where the model is too strict or too loose.

• Overrides need a written reason
• The AI verdict is never overwritten

Keep the record

Every inspection is saved with the photo, heatmap, lot, serial, operator and model version. The dashboard shows defect rate, false rejects and the cost of scrap. Export to CSV whenever a customer or auditor asks.

• Per-part record with photo
• Date-range reports and CSV export