No defective bottle reaches the case

The problem

At full line speed, the eye gives up before the shift ends.

100% human visual control is impossible at real line speed; sample-based inspection leaves gaps, and a batch with a defect can reach distribution undetected.

• Low fill level, a cork sitting proud or crooked, a capsule badly crimped, a back-label skewed or missing: each one is a bottle the end customer will notice on the shelf.
• Checking every bottle by eye at real bottling speed is impossible. Inspection becomes sampling, and sampling leaves gaps by definition.
• A batch with a recurring defect can reach the distributor before anyone at the winery sees it, because the samples happened to be good ones.
• Then the complaint comes back from the importer, with the cost of a return and a damaged relationship attached. And nobody can show which bottles, of which hour, carried the defect, so the whole batch falls under suspicion.

How it fits the IRIS system

Edge — local vision on the bottling line, no image leaves the winery.

Edge — local CNN vision, no image sent to the cloud. Model trained on examples of the specific bottle format.

The model is trained on your bottle, your cork and your label, not on a generic catalog. That is what keeps false rejects low enough for the line team to trust the ejector instead of switching it off in the second week. It is also what makes the ejected bottles useful: each one comes with the image that explains why it left the line, so the filler, the corker or the labeler can be adjusted at the source.

See the full IRIS architecture →

Before and after

Sampled bottles versus every bottle inspected

Aspect	Visual sampling	With iLEAN Edge
Share of bottles checked	1-2%	100%
Fill level, cork and label	Checked when someone looks	Checked on every bottle in milliseconds
A defective bottle	Boxed with the rest	Ejected before casing
Images of the line	—	Processed locally, never sent to the cloud
Distributor and importer complaints	Recurring	Drastic reduction
Format change	New eyes to train	Model switched to the active format

1-2% sampling → 100% bottle control; distributor/importer complaints → drastic reduction.

Impact estimate

Impact estimate on the line — to be validated with your current defect ratio.

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 current defect ratio of your line.
• From 1-2% sampling to 100% bottle control, at real bottling speed. Every bottle, not a sample, has a recorded verdict. The line keeps its speed; the camera adapts to it, not the other way around.
• Distributor and importer complaints drop drastically, and with them the cost of returns and replacements. Fewer complaints also means fewer hours spent by quality answering them.
• And every ejected bottle keeps its image, which explains the defect instead of just counting it.

estimated payback of 5-12 months depending on the current defect ratio. Estimate to validate.

And the fair question from the production manager

"What if it starts rejecting good bottles and slows the line?" — the false reject is the real risk of any vision system, which is why the model learns from good and bad examples of your specific format and under that station's lighting. Classifying a known defect on a known bottle is an anchored task, where the best models drop below 1.5% error [1], and borderline cases go to a person instead of being thrown away.

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