Zero defect at the box-build station

The problem

Box-build is the last station and the least inspected.

That unit reaches the OEM customer and turns into a return, a quality claim, or in the worst case a field investigation.

• Final assembly is manual by nature: connectors seated by hand, screws torqued, harnesses routed, labels applied, enclosure closed. It is where board work turns into a product.
• And it is inspected by sampling, typically 1 to 5 % of units, because inspecting every unit by eye at cadence is not realistic and nobody pretends otherwise.
• So a connector that looks seated but is not, a missing screw on a panel, or a label applied to the wrong face leaves the plant inside a sealed box, and the enclosure makes the defect unreachable from that point on.
• That unit reaches the OEM customer and becomes a return, a quality claim, or in the worst case a field investigation on a program you assemble for several customers at once. In an EMS business the claim costs more than the unit ever did, because what it really costs is scorecard and the next award.

How it fits the IRIS system

Edge — a camera over the station and a model trained on that program's own product.

Edge: an industrial camera over the final assembly station + a CNN trained on good/bad examples of each program's specific product. Millisecond inference, no image sent to the cloud. The defective unit is stopped before packaging.

Inference happens in milliseconds on the station itself: no image leaves the plant, which is what makes the case approvable when the product belongs to somebody else's brand. The defective unit is stopped before packaging, not after shipping — and the operator gets the signal at their own bench, while the unit is still open in front of them.

See the full IRIS architecture →

Before and after

Today's final inspection versus Edge vision

Aspect	Sampling inspection	With iLEAN Edge
Coverage	1-5 % of units	100 % unit control
Unseated connector	Looks fine to the eye	Classified unit by unit
Missing screw or label	Caught if it is in the sample	Caught before packaging
Where the defect is found	At the OEM customer	At the station
Customer claims over visual defects	Recurring	Drastically reduced
Images of the customer's product	—	Never leave the plant

1-5% sampling → 100% unit control. Customer claims over visual defects → drastic reduction.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-12 months, depending on your volume and your current claim rate.
• From 1-5 % sampling to 100 % unit control on the covered product families.
• Customer claims over visual defects drastically reduced — and in an EMS plant a claim costs far more than the unit, because what it really costs is position on the customer's supplier scorecard.
• And the defective unit is stopped before packaging, which removes return freight, rework at the customer's site and the credit note that follows — along with the corrective action report somebody would have had to write.

Payback 5-12 months · from sampling to 100% control. Estimate to be validated.

And the fair question from the production manager

“What if it rejects good units?” — the false positive is the real risk of any vision system, which is why the model is trained on good and bad examples of that specific program's product and under that station's lighting, not on a generic model. Classifying against trained references is an anchored task where the best models drop below 1.5% error [1], and borderline units are escalated to the operator rather than rejected outright.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.