Edge vision on final assembly

The problem

Some defects pass the electrical bench and fail in the vehicle.

- A misseated connector, an incorrectly torqued screw, or a cold solder joint are subtle visual defects that can pass a bench electrical test. - They're also the most expensive defects if they reach a vehicle on the road, triggering an OEM customer complaint.

• At final assembly cadence the human eye tires, and by the end of the shift a subtle defect is easier to miss than at the start. Sampling helps, but it only sees the units it picks.
• A misseated connector, an incorrectly torqued screw or a cold solder joint are subtle visual defects, and they can pass a bench electrical test because contact is still made on the bench.
• Vibration and thermal cycling in the engine bay are what open them up, weeks or months later, when the unit is already fitted to a vehicle.
• They are the most expensive defects once they reach a vehicle on the road: an OEM customer complaint, a warranty return and, if the pattern repeats, a recall discussion.

How it fits the IRIS system

Edge vision over the line — a CNN trained on your own alternators, starters and connectors.

Edge camera over the line with a CNN trained on good and bad images of the specific product (alternator, starter motor, ECU connector). Millisecond inference, rejecting the defective unit before packaging.

The model is trained on good and bad images of your specific product, not on a generic catalog. It runs on the line, in milliseconds, so the defective unit is diverted before packaging instead of being found by the OEM. Inspectors move from hunting every unit to deciding on the doubtful ones.

See the full IRIS architecture →

Before and after

Final inspection today versus Edge inspection

Aspect	Today	With iLEAN Edge
False negatives at cadence	Rise when the line speeds up	Independent of line speed
Misseated connector	Can pass the bench test	Detected and the unit diverted
Screw torqued wrong or missing	Caught only if someone looks	Checked on every unit
Cold solder joint on a visible terminal	Hard to judge by eye	Classified from the image
Evidence per unit	None	Photo linked to the serial number
Warranty return analysis	Starts without images	Starts with the unit's photo at shipment

Human visual inspection with fatigue and false negatives → constant Edge inspection, no fatigue, with photographic evidence per unit.

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 5-10 months, depending on line volume.
• Estimated scrap and rework reduction of at least 30% at the control point, because defects are caught where they are still cheap to fix.
• From human inspection with fatigue and false negatives to constant Edge inspection, with photographic evidence per unit.
• And a lower risk of the defect that costs most: the one that passes the bench and fails in the engine bay, where a single escape can open a warranty campaign.

Estimated scrap/rework reduction of at least 30% at the control point, payback 5-10 months depending on line volume. Estimate to be validated.

And the fair question from the production manager

“What if the camera rejects good alternators?” — classifying known defect types on a fixed product under controlled lighting is an anchored task, where the best models stay below 1.5% [1] error. Doubtful units are not scrapped automatically: they are diverted for a person to decide, and every decision feeds back into the model. Over time the doubtful share shrinks, because the model learns from the calls your inspectors make.

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