Defects caught on the fly

The problem

At cycle speed, the eye is the weakest control in the plant.

100% human visual inspection is impossible at real cycle-time speed, and sampling leaves gaps exactly on the intermittent defects. On a visible Class-A or functional part, that defect reaches the customer: returns, quality claims, risk of losing preferred-supplier status.

• An injection press does not wait: parts come out at the cycle the process needs, and the inspector is the same person at hour eight of a night shift as at hour one.
• 100% visual inspection is impossible at that pace, so the plant samples — and sampling leaves gaps exactly where the intermittent defect lives.
• Flash from a worn parting line, a sink mark from a holding-time drift or a short shot on one cavity can all come and go within a lot.
• On a Class A or a functional part, that defect reaches the customer: returns, a quality claim, and the preferred-supplier status that decides the next award.

How it fits the IRIS system

Edge — a local CNN over the cell, inference in milliseconds, images never leave the plant.

Edge — local CNN vision, no images sent to the cloud. Model trained on good and bad examples of the customer's specific component.

The model is trained on good and bad examples of your specific part, under that cell's lighting, not on a generic defect catalog. Nothing is sent to the cloud: inference happens on the unit at the cell, which is what keeps it inside the cycle and inside your customer's confidentiality clause. That is the difference between a demo and something the quality lead is willing to put between the press and the box.

See the full IRIS architecture →

Before and after

Sampling inspection versus Edge vision on molded parts

Aspect	Today	With iLEAN Edge
Parts inspected	A 1-5% sample	Every part, every cavity
Flash on a worn parting line	Noticed when a customer complains	Flagged on the first parts
Short shot on one cavity	Slips through between samples	Pulled before packaging
The defective part	Boxed with the good ones	Diverted at the reject gate
Criterion at hour eight of the night shift	The same eyes, more tired	The criterion does not tire
Visual-defect PPM at the customer	The number argued over monthly	Driven down and traceable

1-5% sampling → 100% part inspection. Customer visual-defect claims → 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 current scrap and claim ratio.
• From a 1-5% sample to 100% of parts inspected, which is the only way an intermittent defect stops being invisible.
• A drastic reduction in visual-defect claims, and with them the containment costs that follow each one.
• Defect attribution per cavity, which turns a scrap number into a specific mold maintenance job.

estimated payback 5-12 months depending on the 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, and it is handled by training on good and bad parts of your own references, with that cell's lighting. Borderline parts are not scrapped: they are routed to a person to decide, and every decision feeds back into the model. This is not an anchored extraction task like the Connect cases, where the best models drop below 1.5% error [1]; here the criterion is yours and it is learned from your parts.

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