Every bottle inspected before it is cased

The problem

At line speed, the human eye lets through exactly what the premium distributor notices.

Control is visual and human, done by a person who has spent hours looking at identical bottles. The defects that get through are precisely the subtle ones — and those are exactly the ones a premium distributor does notice and does return. In contract production the problem scales: the cosmetic defect is claimed by the client who commissioned the brand, with a contract behind it. That is not a credit note, it is an uncomfortable conversation with whoever gives you volume. And there is a third layer: the tax stamp. A missing or misplaced stamp is not a cosmetic defect, it is a compliance problem.

• In a premium tequila the bottle is part of the product: the distributor and the consumer are paying for the glass, the label and the capsule too.
• Control is visual and human, done by a person who has spent hours looking at identical bottles. The defects that get through are the subtle ones — the slightly low fill, the label one degree crooked, the badly heat-sealed capsule — and those are the ones a premium distributor returns.
• In contract production the cosmetic defect is claimed by the client who commissioned the brand, with a contract behind it. That is not a credit note; it is an uncomfortable conversation with whoever gives you volume.
• And the marbete: a missing or lifting tax stamp is not a cosmetic defect, it is a compliance problem.

How it fits the IRIS system

Edge — AI vision on the line, local inference in milliseconds.

Edge, AI vision on the line. The flow:

The camera inspects 100% of bottles and stores the image of every reject as evidence tied to the batch. The person stays in command: they decide what is reworked and what is scrapped, and review edge cases from a screen of thumbnails instead of staring at the conveyor. Rejects are not lost in a bin: each one has its image, its batch and its reason, and the same images feed the evidence pack of the batch.

• An industrial camera mounted over the line, after filling, labeling and stamp application, and before the case packer.
• Local inference in the plant itself, in milliseconds: no dependence on the cloud or on connectivity, which in a mountain plant is no small detail.
• The network is trained on images of this plant's actual bottles and labels, brand by brand.
• It detects out-of-range fill level, crooked or swapped front and back labels, incomplete capsule or seal, missing or misplaced tax stamp, and visible particles or haze.
• It rejects the bottle before casing and stores the image as evidence tied to the batch.

The person stays in command: they decide what is reworked and what is scrapped, and review edge cases from a screen of thumbnails.

See the full IRIS architecture →

Before and after

Inspection by eye versus Edge vision

Aspect	Visual inspection	With iLEAN Edge
Coverage	Sampling, depends on fatigue	100% of bottles, constant
Fill level slightly low	Passes at line speed	Rejected before casing
Front and back label swapped or crooked	Noticed by the distributor	Detected in the plant
Missing or misplaced marbete	A compliance problem in the market	Caught before the case packer
Evidence per batch	None	Image of every reject
Connectivity needed	—	None: inference is local

From sampling by eye to inspecting 100% of bottles, with visual evidence per batch. From distributor returns and cosmetic claims to detection inside the plant.

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 reduction in escaped defects of at least 30% versus human control at line speed.
• Estimated payback 6-12 months, depending on bottle volume and the average cost of a return.
• Distributor returns and cosmetic claims turned into detection inside the plant, with the image of every reject to show a brand client.
• And a compliance layer on the marbete that today rests on whoever happens to be looking at the moment the stamp lifts at a corner, bottle after bottle.

Estimated reduction in escaped defects of at least 30% versus human control at line speed. Estimated payback 6-12 months depending on bottle volume and the average cost of a return. *Estimate to be validated.*

And the fair question from the production manager

"What if it rejects good bottles?" — the false positive is the real risk of any vision system, which is why the network is trained on images of this plant's actual bottles and labels, brand by brand, under that line's lighting, not on a generic model. In tasks anchored that way the best models drop below 1.5% error [1]. Borderline cases are not simply rejected: they go to the thumbnail screen for a person to decide, and every decision feeds back into training. Thresholds are tightened or loosened by quality, not by the model.

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