Filatura control for PDO Mozzarella di Bufala Campana — the filatore's hands, assisted by vision.

The problem

The filatore's eye, thirty seconds, the result of the whole day.

PDO Mozzarella di Bufala Campana is made in three critical steps: curd acidification, filatura in hot whey, and forming and cooling. The step that decides the final texture is the second one: the curd is immersed in whey at around 90 °C, the filatore stretches it with his hands and, when it "pulls" with the right elasticity, it is formed. Thirty seconds, one decision, the quality of the day.

In most caseifici that lives in three separate places:

1. The filatura vat — whey temperature, often on an old PLC with a panel the filatore glances at out of the corner of his eye.
2. Curd pH before stretching — measured by hand by the quality manager, written down on a sheet.
3. Weight per portion at the outfeed — checked by scale and by eye, with manual sampling.

The expert filatore keeps it all in his head. The problem starts when he retires, when he has an off day, or when the caseificio has to double output and junior filatori come in. The veteran's knowledge lives inside one person — and that is the biggest island in the whole factory.

How it fits the IRIS system

iLEAN Edge does not replace the filatore — it captures his pattern as a permanent capability of the caseificio.

The problem with filatura is not a lack of information: it is information that lives inside one person and is only cross-referenced with the vat when that person is standing in front of it. iLEAN acts as the putty that fills that gap, without asking you to change the vat, the panel or the forming machine. The filatore is still the filatore — Edge gives him a safety net.

Edge sees the filatura vat. Connect captures the pH and the records. The agent cross-references them with the learned pattern and warns the filatore at the line. The person signs — never the other way round.

The iLEAN pieces applied to the filatura of PDO Mozzarella di Bufala Campana:

• Edge — a terminal with machine vision (CNN) and a temperature sensor above the filatura vat. It reads whey temperature, the visible elasticity of the curd and the weight per portion at the outfeed in milliseconds. If something falls outside the range, it fires an actuator (light stack, rejector) and warns the filatore. It works with no network.
• Connect — captures the pH measured by the quality manager (a photo of the meter or a voice reading), the vat records and the outfeed reports. It also captures what comes in from outside (a new Consorzio rule, an importer complaint) at second zero.
• Agent — learns the pattern of a good filatura from the expert filatore, stores it as a reference and applies it to assist the rest of the team. When a vat drifts from the pattern, it does not send an email at 10 p.m.: it warns the filatore in the moment, with the context. The filatore decides.

See the full IRIS architecture →

Before and after

Manual filatura vs. filatura assisted by iLEAN Edge

Aspect	Filatore + panel + sampling	With iLEAN Edge + Connect + Agent
Whey temperature	PLC panel, a glance out of the corner of the eye	Reading at the line + alert to the filatore
Curd elasticity	The filatore's eye	Pattern learned from the filatore, verified by vision
pH before stretching	Manual sampling, handwritten note	Captured by Connect, cross-referenced by the agent
Weight per portion	Scale + eye, sampling	Vision + scale on 100% of output
The veteran's know-how	Leaves with the filatore the day he retires	Captured as a permanent pattern of the caseificio
Batch dossier for the PDO / importer	Rebuilt by hand, days	Ready, with a photo of the vat, in hours

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

• PDO Mozzarella di Bufala Campana caseificio, ~2-5 t/day of pasta filata, 1-2 forming lines, multi-SKU (bocconcini, ovoline, treccia).
• Edge pilot on one filatura vat and the forming machine, with pH capture through Connect and a pattern learned from the expert filatore. First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the frequency of filatura non-conformities and the average cost of a deviated batch or of off-weight portions.
• Expected reduction in scrap from deviated filatura and off-weight portions of ≥ 30%, with a batch dossier for the PDO and the importer.
• The hard lever is the day the expert filatore is not there: the learned pattern lets you run with stable quality even when the expert is away.

And the filatore's reasonable doubt

"What if the AI misreads the elasticity and lets a badly stretched curd through?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply recontextualizes a sensory data point (checking temperature, comparing texture against a pattern), the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN warns the filatore and the person signs. The three safety rings exist precisely for this.

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