Defect caught on the fly

The problem

At 15,000 units an hour, inspection by eye is a sampling plan.

100% human visual inspection is impossible at the real cadence of the most modern lines. Sampling-based inspection leaves gaps that translate into retail returns and lost customer trust.

• The sector's fastest lines reach 15,000 units per hour. Piece-by-piece human inspection at that cadence is mathematically impossible, however good the team is.
• What exists is sampling: one or two percent of loaves or buns checked by eye, usually at the spiral cooler exit or before the slicer. The rest is trusted to the process being stable.
• Everything in between ships: a collapsed top, a burnt crust, a torn bun, a loaf stuck to its neighbor, a pale batch from an oven zone running cold.
• Those gaps come back days later as route returns and complaints from the retail buyer, and with them a loss of trust that no sample plan recovers.

How it fits the IRIS system

Edge — local CNN vision above the belt, inference in milliseconds, ejection before packaging.

Edge — local CNN vision, no need to send images to the cloud (latency, cost, bandwidth). Model trained on the specific format of each line (sliced bread, burger or hot-dog buns).

The camera does not get tired at hour seven and does not sample. Every loaf is classified before the bagger, and the defective one leaves the line on its own, so what reaches the retail shelf is what quality would have approved by hand. Images never leave the plant: inference runs next to the belt, with no latency, bandwidth or cloud cost to justify.

See the full IRIS architecture →

Before and after

Sampling by eye versus Edge inspection at full cadence

Aspect	Today	With iLEAN Edge
Share of production inspected	1-2% sampling	100% of pieces
Collapsed top or burnt crust	Ships unless sampled	Ejected before the bagger
Burger bun versus hot-dog bun	Same tired eye for both	A model trained per format
Images and cloud traffic	—	Processed on site, nothing uploaded
Route returns for visual defects	A recurring cost	Drastically reduced
Defect history per batch	None	Counted and stored

Impact estimate

Impact estimate — to be validated with your bakery's 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 ratio.
• Plus the reduction in retail returns for visual defects, which today reach the customer before they reach quality and cost the freight, the credit note and the buyer's patience.
• From one or two percent sampling to 100% piece inspection, at the line's real cadence and on every shift, including nights and weekends.
• And a defect count per batch that tells the oven and proofing teams where the problem starts, so the cause is corrected upstream instead of rejected downstream.

Estimated payback of 5 to 12 months depending on current scrap ratio, plus reduced retail returns. Estimate to validate.

And the fair question from the production manager

“What if it ejects good bread?” — 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 formats and with that line's lighting. Classifying a known defect on a known format is an anchored task, where the best models drop below 1.5% error [1], and borderline pieces are sent to a review lane for a person to decide. Every decision feeds back into training, so the line's own judgment ends up inside the model.

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