Snack seasoning control with AI — the right dose in every bag, not just in the recipe.

The problem

The recipe says one thing, the drum does another, and the customer can taste it.

On a snack line, seasoning is not applied as an exact weight on each chip: it is applied as a continuous flow of powder onto a rotating drum with a bed of product inside. And between the recipe and the consumer's mouth, half a dozen variables creep in that nobody ever wrote into any system:

• The day's humidity — on a dry day the seasoning flies; on a humid day it compacts and the dispenser discharges differently at the same setpoint.
• The drum fill level — if the upstream line slows down, the seasoning-to-product ratio shoots up.
• The upstream oil sprayer — if the oil cools or the filter clogs, the seasoning does not adhere and ends up on the floor instead of on the chip.
• The particle size of the seasoning batch — a change of supplier or batch and the flow changes without the recipe reflecting it.

The classic system is an hourly sample to the lab, a sensory test and the sales curve. It works 99% of the time. That 1% is the retailer complaint about bags as salty as the sea, or the entire batch sent to rework because the tasting panel downgraded it three hours after packing. The expensive part is not the salt: it is everything that gets labeled, packed, palletized and shipped before anyone finds out.

How it fits the IRIS system

iLEAN does not replace the lab — it puts a continuous signal between two analyses.

The seasoning-control problem is not a lack of recipe: it is information living in islands that, at the moment things drift (SKU changeover, new seasoning batch, upstream stoppage), does not reach the decision-maker in time. iLEAN acts as the putty that fills the gaps between the recipe, the dispenser and the line, without asking you to change the drum or the packing.

Edge sees the coverage on the drum. Connect reads the dispenser, whether it comes from a modern PLC or an analog panel. The agent cross-checks against the SKU recipe and, if the deviation persists, holds product. The person signs — the line does not restart on its own.

The three iLEAN pieces applied to snack seasoning control:

• Edge — a machine-vision terminal (CNN) over the drum outlet. It reads the visible coverage of the product, measures distribution, detects pale zones and clumps. If coverage falls outside the SKU's window, it triggers an actuator (stack light, diversion to a rework bin) within milliseconds. It works without a network: if the plant loses WiFi, Edge keeps reading and holding, because what is critical cannot depend on connectivity.
• Connect — captures the seasoning dispenser reading in whatever format it comes: modern PLC, old isolated local computer, photographed analog panel. And it also captures what arrives from outside: the supplier's email with the new batch certificate, the shift lead's WhatsApp with the SKU changeover moved forward. Everything comes in at second zero, without anyone forwarding anything.
• Agent — cross-checks the visible coverage, the dispenser's actual flow rate, the SKU recipe, the recorded ambient humidity, the line's history and the upstream rate. If the deviation persists over several drum turns, it does not send an email at 10 p.m.: it holds the suspect bin and alerts the quality manager on whatever channel they use. The person validates and signs.

See the full IRIS architecture →

Before and after

Manual seasoning control vs. cross-checked control with iLEAN

Aspect	Hourly sampling + lab	With iLEAN Edge + Connect + Agent
Control frequency	Every 60-90 min, spot sample	Every turn of the drum, continuously
Dispenser reading	Written down by hand or ignored	Captured from the panel or the PLC at second zero
SKU changeover	New recipe, wait for the first analysis	Edge validates coverage from the first turn
Seasoning batch change	Assumed "equivalent" until something fails	Change captured by Connect, agent recalibrates the window
Deviation detection	3 hours later, with product already packed	In line, before packing
File for the IFS/BRC auditor	Rebuilt by hand per batch	Per-batch dossier with coverage, flow rate and signature

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.

• Fried or extruded snack plant, one drum seasoning line, 6-10 SKUs with different seasoning profiles (classic, spicy, BBQ, cheese).
• Edge pilot on one line (camera over the drum outlet + integration with the dispenser + cross-check against the MES recipe). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on the current average cost of a batch sent to rework and the frequency of retailer complaints about inconsistent flavor.
• Expected reduction in scrap from out-of-spec seasoning of ≥30%. The hard lever is what today goes into the rework sack or gets sold as lower-margin private label.

And the quality manager's reasonable doubt

"What if the AI decides to hold a good batch?" — hallucination is a problem of free generation, not of anchored tasks. When the AI merely reads an image of the drum and compares it against the recipe and the measured flow rate, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds 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.