NO BADLY PRESSED CUBE GETS THROUGH

The problem

At line speed, sampling is an act of faith.

On a high-speed cube press or sachet filler, human visual inspection cannot sustain 100% coverage. Sampling leaves gaps, and deformed or badly pressed cubes, torn wrappers, and sachets with sealing defects reach customers — generating complaints and full-pallet returns.

• A cube press and a sachet filler running at commercial cadence produce more pieces per minute than any pair of eyes can classify. The eye is not the problem; the arithmetic is.
• Sampling covers 1 to 2% of the run, which means the defect distribution is inferred rather than observed. Everything you believe about your own defect rate comes from that slice.
• Deformed or badly pressed cubes, torn wrappers and sachets with poor seal integrity leave the plant inside sealed cartons, indistinguishable from the rest of the pallet.
• What comes back is not a piece: it is a complaint and a full pallet returned, with the whole run under suspicion and the distributor asking questions.

How it fits the IRIS system

Edge — every piece classified in milliseconds, on the machine itself.

Edge places a camera over the press or sealer, with a CNN trained on the specific product and format, inferring in milliseconds without sending anything to the cloud, and rejecting the defective piece before cartoning.

Inference runs locally, on the line, with nothing leaving the plant network. A camera over the press is not a data project: the defective piece is ejected before the carton is formed, and that decision never depends on a connection, a cloud service or somebody being awake at three in the morning.

See the full IRIS architecture →

Before and after

Sampling inspection versus Edge inspection

Aspect	Today	With iLEAN Edge
Coverage of the run	1-2%, by sampling	Every piece, at cadence
A crumbling or badly pressed cube	Passes unless somebody is looking	Ejected before cartoning
A sachet seal that is weak but closed	Found by the consumer	Classified at the sealing jaw
Where the defect is caught	In a complaint	On the belt
Full pallets coming back	A recurring cost	Cut off at source
Several formats on one line	Each one hides differently	One model per active format

1-2% sampling → 100% piece inspection. Pressing/sealing complaints → drastic reduction.

Impact estimate

Impact estimate — to be validated against your current reject rate.

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 the rejection rate you run today and on how many formats share the line.
• Complaints about pressing and sealing fall sharply, and with them the returned pallets that carry freight, rework and an awkward conversation with the distributor.
• Coverage goes from 1-2% of pieces to all of them, which changes what you can honestly tell a distributor when they ask how the product is controlled.
• And the defect distribution stops being an inference drawn from a sample and becomes a measurement per shift, per format and per press.

Estimated payback of 5 to 12 months depending on the current rejection rate, with fewer customer complaints. Estimate to be validated.

And the fair question from the production manager

"What if it starts throwing away good cubes?" — 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, under that line's lighting, rather than on a generic defect library. Borderline pieces are escalated for a person to judge instead of being rejected blindly, and every judgment goes back into training. On the classification itself, an anchored task with a fixed frame and a short defect catalog, the best models stay under 1.5% error [1]. And the ejection rate is monitored from the first day, so a model that starts being too strict is visible before it costs you product.

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