A ceramic tile that warps in the kiln is scrap — and the blame lies with the curve nobody cross-checked against the thermocouples.

The problem

The kiln runs fine until a new format changes the stresses — and nobody was cross-checking the panel curve against the thermocouples.

A Sacmi ceramic kiln is one of those assets that costs hundreds of thousands of euros, lasts for decades and runs on an analog panel the firing manager reads at a glance. As long as the catalogue is stable and the curve is dialled in, the line produces good tile. The day a new size, a different glaze or a body with another expansion coefficient comes in, the balance breaks — and the tile leaves the kiln with internal stress that shows up later (when it hits the water in the polisher, when it is bevelled) as edge curl, crowning or a dished centre.

1. The kiln curve and the thermocouples live in different systems. The kiln panel belongs to the firing manager; the thermocouples sit in the SCADA or in an add-on IoT layer. Nobody cross-checks them in real time.
2. The defect only shows up later. By the time the tile reaches the polisher, the kiln load is already fired. A large-format tile cannot be reworked: it is scrap.
3. The firing manager's knowledge walks out with him. He knows how to read the panel and anticipate — but that know-how never enters any system, and the day he retires, half the plant leaves with him.

This is not a kiln retrofit problem. Integrating the panel was always more expensive than the data. That has expired. AI changes the arithmetic: today the panel is read the way a person reads it and cross-checked against the thermocouples without touching the kiln control.

How it fits into the IRIS system

iLEAN does not ask you to retrofit the kiln — it puts a camera in front of the panel and cross-checks the thermocouples you already have.

Connect is the filler between the firing manager's trade and the loose data in the SCADA. It does not demolish the kiln or replace the panel: it fills the gap the SCADA left exposed. The capture grade is tuned to the specific kiln — manual with a fixed camera if the panel is purely analog, intermediate if it has an isolated digital output that can be read without touching the control.

The camera reads the panel. Connect captures the thermocouples. The agent cross-checks, learns the pattern and flags the deviation with margin. A person signs off the correction — the kiln still belongs to the trade.

The iLEAN pieces applied to the Sacmi ceramic kiln:

• Connect (manual grade) — fixed camera facing the analog panel. Cyclic capture of the curve exactly as the firing manager sees it. If the kiln has a small digital output, however old, Connect reads that too (intermediate grade). No control retrofit required.
• Connect (listening outward) — it hooks whatever falls around the kiln: the firing manager's WhatsApp message announcing the format change, the lab email with the new body composition, the forgotten spreadsheet holding the parameters used the last time with that glaze. Everything enters the system at second zero.
• Agents — they cross-check panel curve, thermocouples, batch in progress and composition. They learn the pattern of the tiles that came out right and of the ones that came out warped. They warn the firing manager with margin — the correction is signed off by a person. And when the batch closes, the evidence pack is already built for the audit and for the lab.

See the full IRIS architecture →

Before and after

Panel alone + loose thermocouples vs. live cross-check with Connect

Aspect	Panel + SCADA, unlinked	With iLEAN Connect + Agents
Reading the kiln panel	Firing manager, at a glance	Fixed camera, cyclic capture
Cross-check with thermocouples	Manual, after the fact	Automatic and live
Detecting the warping	At the polisher, when it is already scrap	Anticipated, with margin to correct
Format change / new body	Trial and error, kiln load by kiln load	Cross-checked against the history of good tile
Isolation of the kiln control	Respected because nothing is integrated	Respected — a camera is not an OT network
Firing manager's knowledge	In his head	Captured as a usable pattern

Impact estimate

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

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

• Ceramic plant with a Sacmi kiln and an analog panel, large formats prone to warping, thermocouples already wired to the SCADA or to IoT.
• Connect pilot (fixed camera facing the panel + thermocouple reading + anticipation agent with a history of good and warped tile). First value expected within a few weeks: the first format change already enters the system with the curve captured.
• Indicative payback between 4 and 9 months, depending on how often kiln loads with documented warping occurred in recent years and the average cost of a scrapped large-format tile.
• Reduction of warping scrap of ≥ 30%. The hard lever is a single kiln load saved — it pays for the pilot, and the brand damage avoided weighs more in the medium term.

And the firing manager's fair objection

"What if the AI misreads the panel and raises an alert when it shouldn't?" — hallucination is a problem of free generation, not of anchored tasks. Reading a curve off the panel and comparing it against a history of patterns is an anchored task: the best models have pushed the error below 1.5%^[1]. And even so, nothing critical is decided alone: the agent warns, it does not act. The curve correction is signed off by the firing manager — the trade still belongs to the trade.

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