Line cleaning between a gluten batch and a gluten-free batch — the invisible residue the quality manager signs blind.

The problem

The cross-contact nobody sees until the batch is already made.

In a breadstick and crispbread factory sharing a line between gluten and gluten-free product, the risk is not in the recipe — it is in what remains on the line when the batch changes:

• Flour stuck in a crack of the belt — invisible from the walkway, perfectly capable of crumbling onto the first gluten-free breadstick passing over it.
• Flour dust settled on a hopper's inner wall — the operator opens, looks with a flashlight, sees nothing odd, signs. The camera would have seen it.
• A bolted joint with trapped crumbs — the typical point the daily cleaning skims over because "flour does not get in there". Until it does.
• A doser mouth or mixer outlet with a thin crust — a cloth cleaning, residue left on the unlit edge.

A single contamination reaching the shelf is an AOECS recall, a regional sanction, a RASFF alert and a gluten-free customer who does not come back. And the "gluten-free" mark is earned in years and lost in a week.

How it fits the IRIS system

iLEAN Edge sees the line after the cleaning — and does not let it start until the person signs.

The cross-contact problem on shared lines is not solved by adding another checklist for the quality manager: it is solved by giving them eyes where today they only have the word of the operator who cleaned. iLEAN Edge is exactly that — machine vision cameras looking at every critical point after the cleaning, comparing against that line's "clean reference state" and leaving the release signature to the person, with point-by-point visual evidence.

Edge looks at the hopper, the belt, the doser, the joint and the forming table after the cleaning. If it sees residue, the line does not start. The person validates and signs — never the other way around.

The iLEAN pieces applied to validating the cleaning between a gluten batch and a gluten-free batch:

• Edge — a machine vision terminal (CNN) at every critical point the quality manager defines: the mixing hopper, the transport belt, the doser, the bolted joint, the forming table, the mixer's outlet. Each camera learns that point's "clean reference state" on that line. After the cleaning, it fires an inspection and compares — if it detects flour residue, crumbs or settled dust, it blocks the startup and records the finding's photo. It works without a network: if the plant loses WiFi, Edge keeps inspecting and blocking, because the critical part cannot depend on connectivity.
• Connect — captures the gluten batch's end event (whether from the ERP, the MES, a manual order from the operator or a mark on the line) and triggers the post-cleaning validation walk. It also captures the quality manager's release signature through whichever channel they use (mobile app, earpiece, tablet), and archives it with photo and timestamp.
• Agent — orchestrates the complete run: it marks the gluten batch's end → reminds the cleaning team of the protocol → on completion, launches the Edge inspection round → if there is residue, holds the line and alerts the quality manager through their channel → when the person signs with the evidence in front of them, it releases the gluten-free batch's startup. And it leaves the per-batch-change dossier ready for IFS/BRC/FSSC 22000 audits.

See the full IRIS architecture →

Before and after

A blind release signature vs. a release with point-by-point visual evidence

Aspect	Classic cleaning validation	With iLEAN Edge on the line
Verification after the cleaning	A by-eye inspection by the operator and the quality manager	An AI vision inspection at every critical point, with an archived photo
The line release signature	Blind — signed without having looked inside every joint	With point-by-point visual evidence in front of the manager
Cross-contact detection	Post-production (an ELISA of the first gluten-free batch)	Pre-startup (residue seen before the first package)
Cleaning validation time	Variable per shift and load — sometimes rushed under pressure	A standard run — Edge does not tire at the end of the shift
Hard-to-reach areas	The assumption that "flour does not get in there"	Explicit coverage of the areas the camera can see, a reinforced manual protocol where it cannot
The file for the IFS/BRC/FSSC auditor	Signed cleaning sheets, reconstruction by hand	A per-batch-change dossier: photo + timestamp per point + the manager's signature

Impact estimate

Impact estimate for your plant — to be validated 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.

• Breadstick/crispbread factory with a line shared between gluten and gluten-free product, 1-2 batch changes a day, 6-10 critical points defined by the cleaning plan.
• Edge pilot on the priority critical points (cameras + a startup-block actuator + integration with the release signature). First value expected within a few weeks: the first post-cleaning inspection with archived visual evidence.
• Indicative payback between 4 and 9 months, depending on the average recall/rework cost documented in your sector and the frequency of cross-contact incidents reported in recent years.
• A reduction of cross-contact incidents ≥ 30% against the baseline (the "before" we measure during the immersion) — a defensible floor, not a shiny ceiling.
• The hard lever is a single AOECS recall avoided: product recovered from the shelf, reverse transport, destruction, a potential sanction and damage to the "gluten-free" mark. A single recall pays for the pilot several times over.

And the fair question from the quality manager

"What if the AI camera errs and lets residue through, or worse, says there is residue when there is none and stops the line for nothing?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI limits itself to comparing the current image with a previously validated "clean reference state" (which is exactly what Edge does here), the best models brought the error below 1.5% ^[1]. And even then, nothing critical is decided alone: Edge holds the line and the person signs, seeing the photo that triggered the hold. iLEAN's three safety rings are there precisely for this — the AI proposes, the quality manager decides.

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