The camera that stops a bad weld

The problem

At cell cadence, nobody inspects one hundred percent of anything.

100% human visual inspection is impossible at real robotic-cell pace. Sampling inspection leaves gaps in safety-relevant structural parts.

• The human eye cannot inspect every weld seam at the pace a robotic cell runs. It is not a training problem or an attitude problem: it is physically impossible.
• So the plant inspects a sample — typically one to five percent — and lives with the arithmetic that follows.
• On a safety-relevant structural part, that gap is not an abstract statistic. The seam that was not looked at is the one that shows up in a claim.
• And the borderline defect — a short seam, a spatter-covered nugget, a geometry that drifted a fraction after a fixture knock — is exactly the one a sample is least likely to catch, because it does not announce itself and it does not repeat on every part. It is also the one most likely to survive final inspection and reach the customer's incoming check.

How it fits the IRIS system

Edge — local inference in milliseconds, before the pallet.

Edge: local CNN vision trained on the specific weld seam and geometry of each part number, with no cloud dependency for inference.

Inference runs locally at the cell, with no cloud dependency: the verdict has to exist before the part moves on, and a network round trip does not fit inside that window. That is why this is Edge and not a service in a datacenter, and it is also why a plant network incident never turns into a line stoppage or a batch released without inspection.

See the full IRIS architecture →

Before and after

Sampling inspection versus Edge vision

Aspect	Sampling inspection	With iLEAN Edge
Coverage	1-5% of parts	100%, part by part
Consistency	Depends on fatigue and shift	Objective and constant at cadence
Weld seam and geometry	Judged by eye	Classified by a trained local model
Where the defect is caught	At final inspection, or at the OEM	Before palletizing
Claims for visual or dimensional defects	Recurring	Drastically reduced
Borderline cases	They slip through	Escalated to a person

From 1-5% sampling to 100% part coverage. Drastic reduction of OEM claims for visual/dimensional defects.

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 current scrap and claim ratio.
• From 1-5% sampling to 100% coverage of the parts in the covered family.
• A drastic reduction of OEM claims for visual and dimensional defects, which is where the relationship actually gets measured.
• And the defective part is rejected before palletizing, instead of being found at final inspection, at the customer's incoming check or, in the worst case, in a sorting campaign at their plant.

Estimated payback 5-12 months depending on current scrap/claim ratio. Estimate to validate.

And the fair question from the production manager

“What if it rejects good parts?” — the false positive is the real risk of any vision system, which is why the model is trained on good and bad seams of your specific part numbers, with that cell's lighting and fixtures, and not on a generic dataset. Classifying a seam against a catalog trained on your own parts is an anchored task, where the best models drop below 1.5% error [1]. And borderline cases are not simply rejected: they are escalated for a person to decide, and every decision feeds back into training.

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