The leak is not discovered in the customer's home

The problem

Sampling detects jaw drift late by definition.

Control today is by sampling, with periodic visual inspection and a destructive test every so many units. Between two samples a whole batch can go through with the crimping jaw out of adjustment. And jaw drift is gradual, so sampling detects it late by definition. The defect, moreover, shows up days later: the leak migrates inside the closed carton and is discovered at the point of sale or in the customer's home. Unsellable product at best; a consumer safety incident and a retailer claim at worst.

• An incomplete crimp is a leak of colorant mass inside the carton. A badly seated cap is a peroxide leak. This is not a cosmetic defect.
• Control today is by sampling, with periodic visual inspection and a destructive test every X units: between two samples a whole batch can run with the crimp jaw out of adjustment.
• Jaw drift is gradual, so sampling always catches it late.
• And the defect shows up days later: the leak migrates inside the sealed carton and is found at the point of sale or in the customer's home.
• Unusable product at best; a consumer safety incident and a retailer complaint at worst.

How it fits the IRIS system

Edge — local vision at the line, with no image sent to the cloud.

Edge. Local CNN vision on a plant GPU — sending no image to the cloud, because of latency, cost and bandwidth — with a model trained on the plant's real formats: crimp geometry, presence and legibility of the code, cap seating, bottle fill level. Inference per unit at real line speed, rejection before cartoning — which is where the defect stops being recoverable — and an automatic alert when the reject rate starts to drift, which is the signal that the jaw needs adjusting before it produces defective units.

Inference resolves in milliseconds on a plant GPU. No image goes to the cloud, and that is not a preference: it is latency at line cadence, traffic cost, and not moving product imagery off the site.

See the full IRIS architecture →

Before and after

Today's sampling versus Edge control

Aspect	Sampling control	With iLEAN Edge
Coverage	1-2 % of units	100 %
Crimp jaw drift	Caught after a batch	Caught in the first few dozen units
Where the leak is found	Point of sale or customer's home	Before the cartoner
Cost of the defective unit	The whole kit, already assembled	The tube, before the kit is built
Destructive test	Every X units	Kept, but no longer the only net
Defect evidence	None	Image tied to batch and shift

Impact estimate

Estimated impact — to validate with your own 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 reject rate.
• The rework cost avoided is that of the complete kit — tube, bottle, carton and components — not just the defective tube.
• Fewer retailer complaints from leaks at the point of sale.
• From 1-2 % sampling to 100 % control of units, at real cadence.

estimated payback of 5 to 12 months depending on the current reject rate and the rework cost of the complete kit — which includes tube, bottle, carton and components, not just the defective tube — plus the reduction in retailer claims. *Estimate to validate*.

And the fair question from the production manager

«What if the camera rejects good units?» — the false positive is the real risk of any vision system, which is why the model is trained on the plant's actual formats — crimp geometry, presence and legibility of the marking — and not on a generic model. Borderline units 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.