One hundred percent seal inspection, not a sample

The problem

At eight thousand packs an hour, the eye does not control the seal. It controls a sample.

The throughput of a modern aseptic filler is far beyond what a person can inspect consistently across a full shift. That is why control is done by periodic sampling, which is statistically sound and completely blind to the sporadic defect. And the sporadic defect is exactly the one that does damage: the reel that starts sealing badly mid-shift, the pack malformed right after a format change, the misaligned cap on one particular run. All of them pass sampling without trouble. The cost does not show up on the line. It shows up when the product is already in another country, with the brand on it and no possibility of quick containment.

• The throughput of a modern aseptic filler is far beyond what a person can inspect consistently across a full shift. So control is done by periodic sampling: statistically sound, and completely blind to the sporadic defect.
• And the sporadic defect is exactly the one that does damage: the reel that starts sealing badly mid-shift, the pack malformed right after a format change, the misaligned cap on one particular run. All of them pass sampling without trouble.
• A defective transverse seal on UHT product is not cosmetic: it is the loss of commercial sterility, which shows up swollen on a shelf in the destination country weeks later.
• The cost does not show up on the line. It shows up when the product is already in another country, with the brand on it and no possibility of quick containment.

How it fits the IRIS system

Edge — one hundred percent inspection in milliseconds, before casing.

Edge, in-line AI vision. The flow:

The value is not only in rejecting: it is that the trend becomes visible before it becomes a problem. Every defect is logged against the batch and the reel, so the reject curve per line climbs on a screen instead of being a hunch that this reel seems odd. Hearing that from the operator is not the same as watching a curve rise ten minutes after the reel change — and being able to stop it before the case packer fills a pallet.

• Industrial camera mounted over the conveyor at the filler outlet, before casing.
• Vision model trained on good and bad packs from that specific line and format — not a generic model: incomplete transverse seal, malformed fold, fill level out of range, misaligned or missing cap, illegible coding.
• Local inference, in milliseconds, with no dependency on the network or the cloud. The line does not stop if the internet does.
• Automatic rejection of the defective pack before casing: it never reaches the pallet.
• Every defect is logged against the batch, so the reject curve per line and per reel is visible in real time.

The value is not only in rejecting: it is that the trend becomes visible before it becomes a problem. Hearing that "this reel seems odd" is not the same as watching a curve climb.

See the full IRIS architecture →

Before and after

Today's inspection versus Edge vision

Aspect	Sampling inspection	With iLEAN Edge
Coverage	A sample per interval	Every pack, at filler speed
Incomplete transverse seal	Passes between samples	Rejected before the case
Fill level and cap	Checked by eye	Classified in milliseconds
Where the defect is found	On a shelf abroad, swollen	At the filler outlet
The reel that starts sealing badly	Suspected	A reject curve per reel, in real time
Dependency on network or cloud	—	None: local inference

from sampling-based inspection, to one hundred percent inspection. From finding the defect in the customer complaint, to rejecting it on the line. From intuition about a reel's quality, to a reject curve per batch.

Impact estimate

Estimated impact — to validate with your own 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 of at least 30 % in pack defect complaints.
• Estimated payback 6-12 months per installed line.
• That timeline drops considerably the moment a single avoided recall of exported product is counted.
• And a reject curve per line and per reel that turns intuition about a reel's quality into a number you can act on mid-shift, before the pallet is built and the container is booked.

estimated reduction of at least 30 % in pack defect complaints, with payback of 6 to 12 months per installed line. That timeline drops considerably once a single avoided recall of exported product is counted. *Estimate to validate.*

And the fair question from the production manager

«What if it rejects good packs?» — the false positive is the real risk of any vision system, which is why the model is trained on good and bad packs from your specific line and format, with that filler's lighting, not on a generic model. Borderline cases are not simply rejected: they are flagged for a person to decide, and every decision feeds back into training. Classifying a seal against a learned reference is an anchored task, where the best models drop below 1.5% error [1].

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