AI control of IQF ice cream — a tunnel at −40 °C does not forgive, and neither does the shelf.

The problem

A thousand pieces a minute, four variables in play, and a tunnel that never stops.

On an IQF line, the shift lead lives between three pieces of data that no system ever serves him together in real time:

1. Tunnel temperature and air flow — in the SCADA, with scheduled defrost cycles that move the real temperature away from the setpoint without warning.
2. The condition of the pieces at the outfeed — visible only by eye, on a belt running at 30 pieces per second. If there is clumping, you find out when the bagging line has already put out twenty bad bags.
3. Glazing thickness and final weight — on the checkweigher and in the bath, each with its own screen, and the operator decides by feel whether the setpoint goes up or down.

The consequence: by the time a deviation is detected, there are already hundreds of kilos of scrap in the bin, or dozens of bags reaching the retailer as a block and coming back as a complaint. The classic system works — but it leaves in the hands of the veteran's eye what AI can sustain without a break, shift after shift.

How it fits the IRIS system

iLEAN does not replace your tunnel or your SCADA — it seals the cracks between the line and the decision.

The problem with IQF is not missing technology: it is that information travels slower than product. By the time someone looks at the screen, the problem piece is already bagged. iLEAN acts as the putty that closes the see-decide-act loop at belt speed, without touching the tunnel or the checkweigher.

Edge sees the outfeed belt in milliseconds. Connect reads the tunnel SCADA and the glazing bath. The agent cross-references and proposes the setpoint adjustment. The shift lead decides and signs.

The three iLEAN pieces applied to an IQF line:

• Edge — a machine vision terminal (CNN) over the tunnel outfeed belt. It identifies clumping, anomalous geometry and out-of-range glazing, and triggers an actuator (stack light, rejector, HMI alert) if the rate exceeds the SKU's threshold. It works with no network. If the plant loses WiFi, Edge keeps classifying.
• Connect — captures the tunnel parameters (temperature, air flow, defrost cycles), the checkweigher reading and the glazing bath temperature, whether they come from a modern SCADA or from the old PLC on an isolated line. And it captures what arrives from outside (an email from the supplier with a change to the bath recipe, a WhatsApp from the maintenance manager about a defrost brought forward) at second zero.
• Agent — cross-references the belt image, the tunnel parameters, the weight and the glazing bath. If clumping goes up, it proposes "lower air flow to X or bring the defrost forward by Y minutes" to the shift lead through whichever channel he uses. The person decides; the setpoint does not move without a signature.

See the full IRIS architecture →

Before and after

Control by the veteran's eye vs. cross-referenced control with iLEAN

Aspect	Control by eye + checklist	With iLEAN Edge + Connect + Agent
Clumping detection	At bagging, with scrap already piled up	On the outfeed belt, in milliseconds
Glazing thickness	By feel, adjusted once per shift	Vision + cross-check with weight, proposal to the lead
Scheduled vs. real defrost	Never cross-referenced with piece quality	The agent cross-references cycle and result
SKU changeover in the tunnel	Conservative setpoint by default	Optimal setpoint per SKU, energy saved
Retailer complaint about a block	Rebuild the batch after the fact	Per-batch dossier with image and trace
The veteran's knowledge	Leaves the day he retires	Captured as a permanent capability

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.

• Plant with 1-2 IQF tunnels, multi-SKU (fruit, vegetables or ice cream), daily recipe and setpoint changes.
• Edge + Connect pilot on one line (camera over the outfeed belt + integration with the tunnel SCADA and the checkweigher). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on your current clumping scrap rate, the cost of complaints about blocks and the energy-saving potential of tuning the setpoint per SKU. A reasonable scrap reduction is in the order of ≥30%, to be refined.
• Hard levers: clumping scrap, tunnel energy savings and fewer retailer complaints.

And the maintenance manager's reasonable doubt

“What if the agent gets the setpoint wrong and burns out a compressor on me?” — the agents have no hands on the critical work order: they propose, the person decides and signs. And in anchored tasks (classifying an image, comparing against the SCADA setpoint), the best models brought error below 1.5% ^[1]. The second safety ring validates every proposal before it is shown to the shift lead; the setpoint never changes without a human signature.

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