Kit inspection with computer vision: 63% fewer complaints

The data layer holds reference photos of parts, their identifiers, and variants. The bill of materials sets requirements per configuration and revision. We match the kit photo to the right specification and reference catalog.

First, we isolate the part in the image. DINOv2 describes its appearance as a feature vector. A vector similarity search finds matching reference parts. These are recognition candidates. OCR and the specification rules verify the variant. The operator resolves any doubts.

Here the bushing looks similar, but its marking shows another variant. Beside it, a washer is missing. The system marks both spots and shows what is required versus what is on the tray.

One more item needs checking: its label is unreadable. The system holds approval and routes the case to an operator. An additional photo makes the marking readable. Any uncertainty must be resolved before the kit is accepted.

The operator replaces the wrong bushing. Next, they add the missing washer. The kit goes back for inspection. Only a new photo and a repeat comparison confirm the match. Closing the issue alone is not enough for approval.

The report links the kit ID, its configuration, photos, and inspection result. A QR code leads back to this document. Assembly and the customer get a shared record of what was checked before handover.

In this deployment, customer complaints dropped by sixty-three percent. The system took over repetitive checks. People could focus on exceptions, root causes of nonconformities, and process improvement.