AI control of snack extrusion — the shape at the die rules, not the recipe in the spreadsheet.

The problem

The extruder is not the one failing — the flour changes, the die fouls up, and nobody sees it in time.

A properly tuned extruder is a wonderfully stable machine. The problem is not the machine: it is the conditions it receives and the speed at which they change:

• A new flour silo — slightly different moisture and particle size; the recipe in the spreadsheet no longer applies.
• A fouled or worn die — the hole geometry shifts by tenths, and expansion collapses.
• A fast SKU changeover — the operator has to adjust temperature, pressure and speed with no stable baseline, and over-corrects.
• A slow temperature drift — five imperceptible minutes, but the oven batch is already coming out off size.

The consequence does not arrive as an extruder alarm; it arrives as a pallet of returned bags three days later, or as an oven that has burned energy cooking scrap. The operator knows it, but he cannot spend his life staring at the belt. The classic system (eye + experience) works 90% of the shift. The other 10% is what the quality committee pays for.

How it fits the IRIS system

iLEAN does not change your extruder — it puts eyes where you could not reach and connects what you had lying loose.

Extrusion is a textbook case for IRIS: there is a point on the line (the die exit) where the defect becomes visible long before it does in the oven or at packing, and there are islands of data (flour silo, extruder PLC, cutter panel, recipe spreadsheet) that nobody joins at the critical moment. iLEAN acts as the putty between those gaps.

Edge sees the snack's shape at the die exit. Connect reads the old extruder and the silo. The agent correlates and proposes the adjustment. The person signs.

The iLEAN pieces applied to extrusion control:

• Edge — a vision terminal (CNN) over the belt at the die exit. It measures shape, length, expansion ratio and color of every snack in milliseconds, and keeps a live histogram. When the distribution drifts, it triggers a stack light and proposes a cause. It works with no network. If the plant loses WiFi, Edge keeps watching and alerting, because what is critical cannot depend on connectivity.
• Connect — captures the moisture and particle size of the flour silo (modern sensor, a read of the local PC or a photo of the panel), the extruder PLC values (temperature, pressure, speed), the cutter speed and the recipe in the process manager's spreadsheet. All of it at second zero, with nobody having to forward anything.
• Agents — correlate the visual drift with the captured variables: "expansion dropped when silo 3 was opened, moisture +0.8%, raise screw speed by 4% and lower zone 3 temperature". It proposes this to the operator through the earpiece. The person validates and applies it; the line never adjusts itself.

See the full IRIS architecture →

Before and after

Classic extrusion vs. extrusion assisted by iLEAN

Aspect	Operator + recipe + eye	With iLEAN Edge + Connect + Agent
Drift detection	5-10 min, by eye on the belt	Seconds, with a live histogram at the exit
Flour silo changeover	Same recipe, no moisture compensation	The agent compensates temperature/speed at the changeover
SKU changeover	Typical over-correction, initial scrap	Recipe loaded + guided fine tuning
Fouled or worn die	Detected via a complaint from packing	Shape distribution shifts → maintenance alert
Operation with no network	n/a	Edge keeps watching on panel power
The veteran's knowledge	Lives in his head, leaves with him	Patterns captured, replicable on another line

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.

• Snack plant with 1-2 twin-screw extruders, multi-SKU (corn curls, sticks, shapes), a tunnel oven and a seasoning + packing line.
• Edge pilot on one line (camera over the belt at the die exit + integration with the PLC and the silo + correlation with the recipe). First value expected within a few weeks.
• Indicative payback between 4 and 9 months. Cooking scrap down ≥30% on problem SKUs; unplanned stoppages from die changes down ≥20%.
• Serial quality in food: demanding standards; in automotive the benchmark is in the order of 25 PPM of out-of-spec parts ^[2]. Moving extrusion toward that PPM logic is the direction of travel.

And the process manager's reasonable doubt

“What if the AI proposes some weird adjustment that ruins the batch?” — the Agents propose, the person decides. That is the principle of the three rings: on a critical work order, the hands are the person's. Hallucination is a problem of free generation; in anchored tasks (measuring shape, reading pressure, correlating), the best models brought error below 1.5% ^[1]. The extruder is still yours — iLEAN just takes the "I'm about to screw this up" out of the operator's shift.

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

^[2] Symestic — PPM standards in automotive serial quality.

Related cluster: pet food extrusion, snack seasoning.