Industrial Cheddar — the belt never stops, and the acidity curve drifts before anyone has time to correct it.

The problem

Four variables moving at once, on a belt that never stops.

Industrial cheddaring is the stage where the curd drains whey, is stacked, turned and restacked to acidify and build texture. On a continuous line, control is hard because four variables move at the same time:

1. The day's starter culture — its real activity changes from one supplier batch to the next.
2. Room temperature and humidity — half a degree moves the pH by several tenths.
3. Stacking weight and the number of turns — they define whey drainage, and they are the levers the shift lead actually has at hand.
4. Dwell time on the belt — set by line speed, and not negotiable without stopping production.

The shift lead adjusts weight and turns from the last hour's history and from spot pH probe readings. When something changes and the curve drifts, they notice it — but often when the curd is already coming off the belt. The result: the batch comes out firm or crumbly and drops to a lower commercial grade, at a noticeably lower price per kg. Or straight to rework. On a line running several tons per shift, those downgraded batches add up fast.

How it fits the IRIS system

iLEAN Edge gives the shift lead the cross-referenced data before the end of the belt.

The problem with continuous cheddaring is not a lack of judgment from the shift lead: it is scattered information that does not arrive cross-referenced in time. SCADA has flows and temperatures, the probes have spot pH readings, the belt has a visual history nobody records, and the day's culture batch is on a delivery note. iLEAN acts as the putty that cross-references those islands on the belt itself, without asking you to change the MES, the SCADA or the line.

Edge sees the belt (stacking, turning, thickness). Connect captures the day's culture and the room conditions. The agent cross-references it with the pH curve and warns before the end of the belt. The shift lead decides and signs.

The iLEAN pieces applied to industrial cheddaring:

• iLEAN Vision (Edge) — a terminal with a camera and a CNN over the cheddaring belt: it identifies stacks, counts turns, measures stack thickness and the apparent texture of the curd. It works with no network: if the plant loses WiFi, Edge keeps reading and recording.
• iLEAN Edge on sensors — captures the continuous pH curve from the inline probes (existing or new), plus room temperature and humidity. It cross-references them with vision without going through the cloud.
• iLEAN Connect — captures the day's culture batch (which supplier, which date, which reported activity) and the recipe of the SKU in progress. It closes the loop between what went onto the belt and what comes off it.
• Agent — cross-references vision + pH + conditions + culture batch. It builds the line's historical pattern: which combination of weight, turns and time keeps the curve inside the optimal window for each SKU. When a new batch comes in off pattern, it warns the shift lead before the end of the belt, with room to adjust weight or turns. The shift lead signs; the agent never touches the parameters on its own.

See the full IRIS architecture →

Before and after

Cheddaring controlled by spot probes vs. cheddaring cross-referenced with iLEAN Edge

Aspect	Classic control	With iLEAN Edge + Connect
Reading stacking and turning	The shift lead's eye, on walk-throughs	Continuous CNN, count and thickness per stack
pH curve	Spot probes, alarm only when out of range	Cross-referenced with vision and culture, batch by batch
Detecting drift	When the curd is already coming out firm	Before the end of the belt, with room to correct
Memory of which culture / which pattern	Lives with the veteran shift lead	Pattern learned batch by batch, reusable
Rework and drops to a lower grade	Frequent on culture changes	Expected reduction ≥ 30% (estimate)
File for an IFS/BRC audit	Rebuilt by hand	Automatic dossier per batch

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

• Continuous industrial Cheddar line running several tons per shift, two or three shifts, multi-SKU (Mild / Mature / Extra Mature) with different pH windows.
• Edge pilot on the cheddaring belt (overhead + side cameras + integration with the pH probes + the room SCADA + the starter culture delivery note). First value expected within a few weeks: a pattern-drift warning before the end of the belt, agreeing with the veteran shift lead's judgment ≥ 90% of the time.
• Expected reduction in batches dropping to a lower grade and in rework ≥ 30%, conservatively. Indicative payback between 4 and 9 months depending on tonnage per shift and the price gap between commercial grades.
• The hard lever is the batch that stays in its commercial grade — a Mature that does not fall to Mild pays for the pilot fast on a large line.

And the shift lead's reasonable doubt

“What if the AI gets the recommendation wrong and costs me a batch?” — the AI does not touch weight or turns. It warns. Hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely cross-references time series (vision, pH, conditions, culture batch) and warns early, the best models brought error below 1.5% ^[1]. And even so: the agent proposes, the shift lead signs. The line does not readjust itself.

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