The defect gets diverted before it reaches packing

The problem

On a registered bioinput, a crooked label is a compliance problem.

At the line's real speed, the human eye tires and lets through bottles with incorrect filling, a poorly closed cap, or a crooked or illegible label. In a bioinput carrying a sanitary registration, an illegible label or a cap that might leak isn't a cosmetic flaw — it's a safety and regulatory compliance problem.

• At the line's real speed, the human eye tires and lets through containers with incorrect fill, a poorly closed cap, or a crooked or illegible label. Late in the shift, after thousands of identical bottles, that is not a matter of diligence.
• On a biopesticide or biostimulant carrying a sanitary registration, an illegible label or a cap that might leak is not cosmetic: it is a safety and regulatory problem.
• Today the check is a 1-2% sample. Everything between samples goes out uninspected, and the gaps are guaranteed. When the defect comes from a capper drifting slowly, the sample catches it hours late.
• A leaking jerrycan of live culture in a distributor's warehouse is a complaint, a return and a question about every other container in that batch, which then has to be checked by hand.

How it fits the IRIS system

Edge — a container-by-container model per format, inferring in milliseconds.

iLEAN Edge places an overhead camera over the line, with a CNN trained on good and bad containers for each concrete format: bottle, drum, jug. It infers in milliseconds per unit.

The model is trained on good and bad containers for each concrete format — bottle, drum, jerrycan — on your own line and lighting, not on a generic catalog of containers. It decides at line speed and diverts the defective unit; it does not stop the line for a borderline case, it escalates it. The operator sees the image and decides in seconds, and that decision trains the next version of the model.

See the full IRIS architecture →

Before and after

Sampling at the bottling line versus Edge inspection

Aspect	Visual sampling	With iLEAN Edge
Containers inspected	1-2%	100%
Cap not fully closed	Found if it happens to be sampled	Detected and diverted
Fill level	Checked by eye on a sample	Checked on every unit
Illegible registration or batch print	Reaches the distributor	Stopped before packing
Format change bottle to jerrycan	Same tired eye	Model switched with the format
Borderline container	Passes	Escalated to a person

Before: 1 to 2% sampling, guaranteed gaps. After: 100% container inspection, diverting the defective one before final packing. Estimated payback 5 to 12 months depending on the current defect ratio.

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 your current defect ratio.
• From spot sampling of 1-2% to 100% container inspection, with the defective unit diverted before final packing and logged against the batch it belongs to.
• Fewer complaints and returns from distributors for leaking caps or unreadable labels on registered products, which are the claims that damage the relationship most.
• And every rejected container is stored with its image, which shows whether the defect comes from the capper, the labeler or the filler, so maintenance fixes the cause instead of the symptom.

Estimated payback 5-12 months · from spot sampling to 100% container inspection.

And the fair question from the production manager

“Won't it reject good bottles and slow the line?” — the false reject is the real risk, which is why the model is trained on your formats and your line's lighting, not on a generic catalog. Classifying a cap, a fill level or a label against a known reference is an anchored task where the best models drop below 1.5% error [1], and borderline units are escalated to a person instead of being thrown away. Every decision the operator takes on those units is fed back into training.

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