Eye formation in Emmental AOP — anticipating the warm room before the defect is born.

The problem

The defect shows up when it can no longer be corrected.

Emmental AOP is one of the most technically demanding cheeses in the world. The AOP specification defines eyes that are bright, regular, from the size of a hazelnut to a walnut (≈1–3 cm), evenly spread through the paste. And a single organism produces them: Propionibacterium freudenreichii, which consumes lactic acid and releases CO2; the CO2 stays trapped in the closed paste and forms the eyes.

There are three ways for a wheel to come out wrong and lose AOP grade:

1. Sparse or blind eyes — the propionic fermentation never started, or was cut short.
2. Eyes that are too large or irregular — the fermentation ran away through excess temperature or time.
3. Cracks (Nachgärung) — a late secondary fermentation that splits the paste open.

The operational problem: classic control is temperature and humidity probes + the master cheesemaker's eye + a sample cut at the end. By the time the master cuts the sample and sees something went wrong, weeks in the warm room and months of ripening have passed; the decision that caused the defect is far behind. AOP does not accept "we'll fix it on the next batch" — this wheel loses several euros a kilo. On an 80–100 kg wheel, that is thousands of euros.

How it fits the IRIS system

iLEAN Edge does not replace the master cheesemaker — it gets the information to him sooner.

The master cheesemaker knows what has to happen in the warm room. What he lacks is early feedback: a signal saying "the propionic fermentation on this run is falling short" or "it is running long" while he can still act on temperature, on airflow or on the exit time. iLEAN acts as the putty between what already exists (probes, culture, curve, sample at the cut) and what was missing (CO2 kinetics measured live, eyes counted and mapped by vision).

Edge measures CO2 in the warm room and learns the kinetics of each run. Vision counts and measures the eyes at the sample cut. The agent cross-references both with the culture and the thermal curve. The master gets the alert days ahead — and signs the correction.

The iLEAN pieces applied to eye formation in Emmental AOP:

• Edge — CO2 sensors in the warm room. They measure the CO2 release kinetics of each run. Trained on your history, the system tells a propionic fermentation that is "below target" from one that is "running normally" and from one that is "about to run away" — days before the result becomes visible at the cut.
• Edge — vision at the sample cut. A CNN counts eyes per unit of surface, measures their size distribution and detects cracks. It builds quantitative feedback for the master: not "it came out fine", but "mean 1.7 cm, σ 0.4, no cracks".
• Agent. Cross-references CO2 + vision + culture + thermal curve + recipe. If the kinetics of a run drift away from the AOP pattern, it proposes the correction to the master (raise or drop 1 ºC, extend or shorten the warm room) with days to spare. The person validates and signs; the system never touches the room on its own.

See the full IRIS architecture →

Before and after

Classic control vs. a warm room assisted by iLEAN Edge

Aspect	Classic control (probes + the eye)	With iLEAN Edge (CO2 + vision + agent)
Feedback on the result	Sample cut, weeks later	CO2 kinetics days earlier + quantitative vision at the cut
Room to correct the warm room	Effectively none	Days — the real time frame of propionic fermentation
Detecting cracks (Nachgärung)	Only at the final cut	An abnormal CO2 signal raises the alert before long ripening
Counting and sizing the eyes	The master's visual estimate	CNN — mean, spread, spatial distribution
Operation with no network	n/a	Edge keeps measuring and alerting on the power it has
The master's knowledge when he retires	It leaves with him	Captured as a reusable pattern for the next generation

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 warm room. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• AOP dairy with 1–2 warm rooms, several hundred wheels a week, documented occasional drift in eye size and distribution.
• Edge pilot in one room (CO2 sensors + a vision head over the sample-cutting table + the agent that learns the kinetics of each culture). First early warning within a few weeks.
• Indicative payback between 4 and 9 months, depending on how many wheels a year currently drop out of AOP grade and on the price gap between AOP and second grade at your retailer.
• The hard lever is every wheel that stays in AOP instead of dropping to second grade — plus capturing the master's knowledge as a reusable pattern for when the next generation walks into the room.

And the master cheesemaker's reasonable doubt

“What if the system gets it wrong and makes me touch a run that was doing fine?” — hallucination is a problem of free generation, not of anchored tasks. Here the AI merely recontextualizes measured signals (CO2, thermal curve, eye count) into a recommendation; the best models brought error below 1.5% ^[1]. And the system never touches the room: it warns you, you decide. The three safety rings exist precisely for this.

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