The Burst Croquette Never Reaches the Bag

The problem

At line speed, a human eye covers one or two percent of what goes into the bag.

At real line speed, 100% human visual control is impossible. With 18 different SKUs, the 'correct shape' pattern also changes from product to product.

• A burst croquette, one with the filling showing through, one with a bare patch where the breadcrumb never took: all of them are obvious on a tray, sitting still, and none of them is reliably caught at cadence on a moving belt.
• With 18 SKUs the 'correct shape' itself changes from product to product: a bite-size appetizer, a 20 g cylinder and a hand-formed artisan piece are not the same reference.
• So visual control ends up being a sample of one or two percent, checked by whoever is free, against a pattern that changed with the last SKU changeover.
• And the defect that gets through does not stop at the bag: in direct sales it arrives at a consumer's home under your own brand, with a photo attached to the review.

How it fits the IRIS system

Edge — inference per piece in milliseconds, with the model of the active SKU.

A camera above the line infers per piece in milliseconds, using a model trained on the specific SKU, with no image ever sent to the cloud. A burst, malformed or incompletely breaded piece is automatically diverted before packaging, without slowing the line's pace.

The image never leaves the line: inference runs locally, so cadence does not depend on the plant network and no picture of your product ever reaches a cloud. Rejecting before packaging also means the film, the tray and the freezing were not spent on a piece that was never going to ship — and that saving is almost always left out of the business case.

See the full IRIS architecture →

Before and after

Today's visual control versus Edge inspection

Aspect	Visual sampling	With iLEAN Edge
Coverage	1-2% sampled	100%, piece by piece
Speed	Slows down to look	Milliseconds, at line pace
A burst or bare piece	Gets through between samples	Diverted before packaging
Pattern per SKU	Changes with each changeover	The model of the active SKU
Defect history	A count on a sheet	Per SKU, shift and run
Images sent to the cloud	—	None: inference is local

1-2% human visual sampling, defects slip through → 100% visual control, in milliseconds per piece.

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 reject ratio and on how much of it is caught today.
• From 1-2% sampling to 100% inspection, without slowing the line's pace by a single piece per minute.
• Fewer visual-defect complaints, which in direct-to-consumer sales arrive as a public review with a photograph attached rather than as a quiet claim from a distributor.
• And a defect history per SKU, shift and run that shows which format, which batter and which filling are actually generating the rejects, instead of a single number argued about in the morning meeting.

Payback 5-12 months · 1-2% sampling → 100% inspection. Estimated payback is 5 to 12 months depending on the current reject ratio, with fewer visual-defect complaints as an added benefit. Estimate to validate.

And the fair question from the production manager

"What if it rejects good pieces?" — the false positive is the real risk of any vision system, which is why the model is trained on good and bad pieces of your own SKUs, under the lighting of that station, and never on a generic model. Borderline pieces are not simply dropped: they are escalated to the operator, and every decision feeds back into training. On the classification itself the task is anchored, with the best models below 1.5% error [1].

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