AI control of the cheese curing chamber — quality stops depending on the shift.

The problem

The BMS shows green, and the batch comes out defective.

Almost every aged-cheese plant has a BMS or a decent climate control system — the problem is rarely that there is no thermostat. The problem is that the BMS controls one point, usually near the return or near the supply of the climate unit, while the chamber as a whole breathes differently:

• The top and the bottom of the chamber sit at different humidity levels.
• The area right by the door takes the hit on every opening — invisible if there is no sensor there.
• A recirculation fan losing airflow because of a slack belt shifts the whole thermal map, and the BMS never notices because its single sensor is in its usual spot.
• At night, with no operators on site, a four-hour drop is discovered at 7 a.m. on the walk-round — far too late for a batch in a critical phase.

The master cheesemaker compensates with his eye and his experience. It works 99% of the time. That 1% is the batch of hundreds of kilos that comes out of spec, taking months of chamber time with it. And when he retires, that 1% goes up.

How it fits the IRIS system

iLEAN does not replace your BMS — it gives it eyes where it could not reach, and an agent that never sleeps.

Your climate control works — the BMS does its job. What is missing is a representative measurement (several points per chamber), operational memory (what happened the last few times this chamber drifted), and an agent that cross-references each batch's curing phase with the real pattern. iLEAN acts as the putty that seals the cracks between the BMS, the master cheesemaker and the shift paperwork — without asking you to change a thing.

Edge measures where the single sensor never reached. The agent predicts the deviation using the batch history. The person signs — the batch is never released on its own. Agents do not sleep; the chamber is never left alone.

The three iLEAN pieces applied to chamber control:

• Edge — a terminal with sensors distributed across levels and zones, a door sensor, and a read of the existing climate system (without replacing it). It works with no network: if the plant loses WiFi, it keeps measuring and recording on panel power.
• Connect — captures everything that happens outside the BMS and reaches the chamber through non-integrated channels: the climate technician writes on WhatsApp that he has changed the fan belt, the production manager dictates by voice that the long door opening at 11:20 was to pull a batch out — all of it enters the system at second zero and stays linked to the chamber and the batch, not buried in a lost email.
• Agent — cross-references the current pattern with the chamber's history and the phase of every batch inside it. It predicts the deviation before the range is breached — not at the threshold, before it. It alerts the on-call manager's channel. The person decides and signs — the agent proposes; it does not open the batch or stop the plant on its own.

See the full IRIS architecture →

Before and after

Single-sensor BMS vs. a chamber read continuously with an agent

Aspect	Classic BMS + walk-round	With iLEAN Edge + Agent
Measurement points per chamber	1 (near the climate unit)	Several, distributed by level/zone
Detection of a night-time deviation	Next morning, on the walk-round	At second zero, to the manager's channel
Operational memory of the chamber	In the master cheesemaker's head	Captured as a pattern, it stays in the plant
Door-opening sensor	No, or only in some chambers	Yes, in every chamber that matters
Deviation prediction	Reactive, on a fixed threshold	Anticipated, by pattern + batch phase
Operation with no network	n/a	Edge keeps measuring on panel power

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.

• Standard curing chamber, existing BMS kept in place, Edge sensing installed in parallel. Pilot in a single chamber.
• First value expected within a few weeks: a real thermal map of the chamber, fine-grained alerts to the manager's channel, a working per-batch dossier.
• Indicative payback between 4 and 9 months, depending on how much scrap from night-time deviations weighs in your product mix and the average cost of a batch ruined by an out-of-range excursion.
• Hard levers: chamber incidents that reach the batch down ≥ 30% (conservative estimate); the night-time "blind spot" eliminated; the master cheesemaker's judgment captured as a persistent pattern.

And the technical director's reasonable doubt

“What if the AI floods me with alarms?” — positive pre-emption: the agents are not fixed-threshold alarm generators. They cross-reference the current reading with the chamber's history, with the curing phase of every batch inside, and with the pattern of whoever is on call. What reaches the manager's phone is the deviation that matters, not the noise. And the system is transparent — the operator can see why the agent raised the alert, and the technical director tunes the sensitivity without touching code. In anchored tasks (reading a real value and comparing it against a pattern), the best models brought error below 1.5% ^[1].

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