THE LINE WON'T START WITHOUT CAMERAS OK

The problem

The cleaning procedure is sound. What is missing is proof it was followed.

Several brands coexist on the same physical line — one shift with gluten or milk, the next without that allergen, or simply a different commercial brand. Today cleaning is signed off with no visual evidence: the responsibility for saying 'it's clean' falls on a human, and residue from the previous recipe can contaminate the first batch of the new format.

• The same physical line runs one shift with gluten, celery or milk in the recipe and the next without it, or simply under a different commercial brand. The equipment does not know the difference.
• Cleaning is signed off with no visual evidence, so the whole weight of saying 'it is clean' rests on one person at the end of a long changeover, usually with the next run already late.
• Residue of the previous recipe in the doser inlet or under the press contaminates the first batch of the new format, and nothing on the line stops it leaving.
• Six months later the same gap becomes a question from an auditor, or from a consumer whose allergen was not on the sachet, and there is no good answer to either.

How it fits the IRIS system

Edge with JIDOKA AI — the line does not start until every camera says OK.

Edge + JIDOKA AI: fixed cameras at the doser, the press/filler and cartoning compare the current state against a 'clean to start' reference state. The line won't start until every camera reports OK and the quality supervisor signs off on visual evidence.

The camera does not replace the person who signs. It gives them something to sign on: each critical point compared against its clean-to-start reference, photographed and dated. What changes is not who decides, but what the decision rests on.

See the full IRIS architecture →

Before and after

The changeover signed blind versus the changeover signed on evidence

Aspect	Today	With iLEAN Edge
What the sign-off rests on	A tick in a box	Each point against its reference image
Residue in the doser inlet	Seen if somebody looks in	Holds the startup
Under the press and in the carton magazine	Out of easy reach	Covered by a fixed camera
First batch of the new brand	Under quiet suspicion	Released against evidence
Proof of it six months later	A signature and nothing else	Dated photographs per point
A week with many changeovers	Risk stacks up unseen	Each one closed with its own proof

Blind sign-off with cross-contamination risk → sign-off based on visual evidence at each critical point.

Impact estimate

Impact estimate — to be validated with your changeover calendar.

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.

• No payback range is claimed here: the return is tied to how many format and brand changeovers you run per week, and inventing months would be dishonest.
• What is protected is the revenue of the affected format, because a contaminated first batch takes the whole run with it, and sometimes the brand's slot on the shelf.
• Cross-contamination between an allergen recipe and the next one stops depending on how tired the changeover crew is, or on how far behind the schedule has slipped.
• And every changeover leaves behind, at no extra cost, exactly the evidence that case 11 turns into an audit dossier for IFS or BRC.

Protects the affected format's revenue; payback depends on the number of format/brand changeovers per week. Estimate to be validated.

And the fair question from the production manager

"Can a camera really tell clean from almost clean?" — the comparison is anchored: same point, same frame, same lighting, against a reference state approved by quality, where the best models stay under 1.5% error [1]. And the system never declares the line clean by itself: it proposes, the quality supervisor signs on the images, and anything ambiguous holds the startup instead of releasing it. The asymmetry is deliberate, because a false hold costs minutes and a false release costs a run.

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