Listeria environmental program in soft cheese — the swab where the risk is, not where the grid says.

The problem

Listeria does not stay still — the grid does.

The classic Listeria monocytogenes environmental program runs on a fixed grid: so many points per room, one round a week, always the same spots. The quality manager designs it with care, HACCP approves it and it is followed to the letter. The problem is that in a soft cheese dairy the risk does not stay still:

1. Condensation moves — it depends on the relative humidity of the room, on how often doors are opened, on the shift change.
2. Traffic moves — the soles coming through change: the route of a new operator is not the route of the veteran who knows the sticky patch by the drain.
3. Biofilm moves — the moment the cleaning schedule slips a few hours or the detergent falls short on one seal, biofilm finds shelter in places you are not looking.

The classic system works 95% of the time. That 5% is the positive that turns up in an area the grid never reached and that turns into a product recall in any Member State. And the committee's question, when RASFF publishes the alert, is not “did you sample?” but “did you sample where the risk was?”.

How it fits the IRIS system

iLEAN Edge does not increase the number of swabs — it puts them where they are needed.

The problem with Listeria sampling is not the number of swabs: it is where they are spent. iLEAN acts as the putty between what the plant shows and what the quality technician decides — without asking you to change the HACCP plan, the lab or the kits.

Edge sees the room. The agent cross-references condensation, traffic, the cleaning schedule and the history of positives, and proposes the week's swab map. The technician adjusts and signs. The swab goes where the risk has moved.

The iLEAN pieces applied to the Listeria environmental program:

• Edge — terminals with machine vision (CNN) in the ripening and packing rooms. They detect condensation, a persistent puddle, a door opened over and over, a cart passing through, boot soles carrying soil into a clean zone. They work without a network: as long as the panel has power, capture continues, because what is critical cannot depend on WiFi.
• Connect — captures the cleaning plan (wherever it lives: an app, a spreadsheet, the hygiene manager's notebook), the lab's history of positives (wherever it lives: a LIMS PDF, an email from the external lab, a photo of the report) and the HACCP list of zones.
• Agent — cross-references what Edge sees with what Connect captures and proposes the week's sampling map, prioritizing areas where the risk has moved. The quality technician reviews, adjusts and signs. The sampling round does not restart on its own.

See the full IRIS architecture →

Before and after

Fixed grid vs. sampling guided by iLEAN

Aspect	Classic environmental program	With iLEAN Edge + Connect + Agent
Plan design	Fixed grid, reviewed once a year	A living plan, adjusted by risk every week
Hot spots detected	Only after a positive	Before sampling, from the image of the plant
History of positives	A PDF in a folder, checked “when there is time”	Cross-referenced automatically with the weekly map
The technician's work	Wheeling the swab cart around	Investigating positives and adjusting the plan
Operation without a network	n/a	Edge keeps capturing on the panel's power
File for IFS/BRC/FSSC	Rebuild the grid and the results	Weekly dossier justifying every single point

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Soft cheese dairy with 1–3 ripening rooms plus packing, a classic environmental program on a fixed grid, several dozen swabs a week.
• Edge pilot in the highest-risk room + Connect over the LIMS history of positives. First value within a few weeks: the first risk-guided weekly map, the first real adjustment to the plan.
• Indicative payback between 4 and 9 months, factoring in the avoided cost of a single recall caused by a positive in product and the qualified time saved for the quality technician.
• Hard lever: a ≥30% reduction in sampling time spent on low-risk areas, redirected to critical areas and to investigating positives. A single avoided recall pays for the pilot.

And the quality manager's reasonable doubt

“What if the AI prioritizes badly and we stop sampling an area where the positive later turns up?” — hallucination is a problem of free generation, not of anchored tasks. When AI merely recontextualizes what it sees (humidity, traffic, condensation) against the real history, the best models brought error below 1.5% ^[1]. And even so, the plan is not decided alone: iLEAN proposes the weekly map and the quality technician signs. The HACCP plan is still the framework; the agent only adjusts within the corridor.

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