A 1998 tunnel oven with an isolated Windows 2000 — and a per-batch baking dossier the IFS auditor asks for tomorrow.

The problem

The oven works perfectly. The problem is what does not come out of it.

In a breadstick and crispbread factory, the tunnel oven is the most expensive piece and the most reliable. It has been baking for 15, 20, sometimes 25 years. The baking curve was tuned back in the day by the master baker — together with the technician of a manufacturer that has since closed — and the oven has delivered ever since. The problem is not the oven; it is what surrounds the oven:

• Only the master baker sees the baking curve. It lives in their head, in a crumpled folder next to the panel, and in what the operator jots down by hand when they remember. The day the master retires, out goes the baking pattern of the thin crispbread, the one for the olive-oil breadstick, and the trick to keep the tunnel's exit from coming out pale when the hall's air changes with October's humidity.
• The industrial PC is isolated from the network. It is a Windows 2000 (sometimes Windows NT) with the oven manufacturer's proprietary driver. It has no usable Ethernet, no Internet, and the IT department has spent years asking "please, do not touch it". They are right: that PC controls the baking, and if it goes down, the line stops.
• The IFS/BRC auditor asks for per-batch baking data and there is no record. There is a paper sheet the shift lead signs at the end of the day. No curve, no setpoint, no recorded alarm. The non-conformity lands in the report and from the next year on the customer demands corrective action — or leaves for the competitor that does have it.

This is not negligence. Integrating old machines was unviable for decades: the project's cost exceeded the saving, the drivers did not exist, and modernizing the whole oven cost more than buying a new one. That equation has changed.

How it fits the IRIS system

Connect, intermediate mode — the putty between your 1998 oven and the rest of the 21st century.

iLEAN Connect was designed with this kind of machine in mind. The philosophy is capture graduation: Connect adapts to what your machine can give, not the other way around. Three modes depending on the oven in front of you:

Connect reads the old PC's serial port or the screen with a camera. The Agents cross the curve with the ERP's batch. The per-batch dossier assembles itself. The master baker stops being the sole keeper of the baking pattern — without losing their authority over the line.

The effect in a breadstick/crispbread factory:

• Manual — a 100% analog panel, needles and pilot lights. A fixed camera points at the panel and a neural network reads the values as the master would. The civil work is a bracket and a power cable.
• Intermediate (your tunnel oven's) — the machine has an isolated, old industrial PC with an RS-232/RS-485 serial port or a flat-file export. Connect connects to that port or reads the export folder, and extracts temperature by zone, belt speed, setpoint, alarms and events. Nothing new is installed inside the old PC — Connect is the external box that reads.
• Integrated — the oven has OPC-UA or modern Modbus. Direct integration to the SCADA or the MES.
• A real baking curve, per batch, in real time. Every tray leaving the tunnel has associated the temperature by zone, the seconds spent in each stretch and the alarms that fired. That is what the IFS auditor asks for, and what never existed on paper.
• The master baker's knowledge is captured as data. Not to replace them — so that the day they are on vacation, the shift's second has the pattern of the last 200 correct batches to guide them. The veteran's eye, at last, stops being an island.
• It works without a network. Connect captures locally and syncs when it can. The plant can lose WiFi for three days: the baking goes on, the capture goes on, the dossier is rebuilt afterwards without gaps.
• Bidirectional but with a brake. Connect transports, the Agents reason, the person signs. The system can suggest to the shift lead that the curve is drifting because the hall's air is more humid than average — but the one who decides to slow the belt is the master, not the agent.

See the full IRIS architecture →

Before and after

A tunnel oven with a notebook vs. a tunnel oven with Connect

Aspect	Today: a paper sheet + an isolated PC	With iLEAN Connect intermediate mode
The batch's baking curve	In the master baker's head	Captured in real time, by zone, by minute
Data for the IFS/BRC auditor	A sheet signed at the end of the shift, no curve	A per-batch dossier with setpoints, the real curve and alarms
The day the master retires	The baking pattern leaves with them	The pattern saved as a reference for the successor
The oven's alarms	A pilot light seen if someone walks by	A timestamped event, notified by channel
The cost of touching the industrial PC	n/a — IT forbids going near it	Zero — Connect reads from outside, nothing installed inside
Integration with ERP/MES/Excel	Manual, weeks after the batch	Automatic cross-check with the work order

Impact estimate

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

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Mid-sized breadstick and crispbread factory, 1-2 tunnel ovens 15-25 years old, an isolated Windows 2000/NT industrial PC or a mixed analog-digital panel.
• Connect pilot on one oven (a Connect box + serial port reading or a camera over the panel + integration with the ERP/Excel batch). First value expected within a few weeks — the per-batch dossier starts generating the day the box is hung.
• Indicative payback between 3 and 9 months, depending on the audit dossier's weight with your end customer (retailer, UK/USA export) and the documented cost of the last baking incident.
• The hard lever is a single IFS/BRC non-conformity avoided in an audit — or a single customer complaint over a badly baked batch that can now be traced to the exact minute. That pays for the pilot.
• A career anchor: the powder coating case documented in iLEAN's base closed its pilot with a real year-1 investment near €30,000 and a conservative payback of ~3 months, with first value in weeks. Same mechanics, another sector: the order of magnitude applies.

And the fair question from the plant director

"What if Connect misreads the panel and records a temperature that is not the real one?" — hallucination is a problem of free generation, not of anchored tasks. Reading a 7-segment display, a thermometer needle or the old PC's flat file and moving that value into a database is an anchored task: there is a source, there is a value that can be verified. In tasks of this kind, the best models brought the error below 1.5% ^[1]. And even then, nothing critical is decided alone: the data is saved, the agent proposes, and the master baker or the shift lead signs. The three safety rings are there precisely for this.

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