The wheat blend, digitized from minute one

The problem

The recipe that defines the flour is written on a sheet nobody can query.

The blend sheet links source silo, ratio, moisture and target protein for the batch. Today it lives on paper in the milling control room and gets transcribed -if at all- hours later, when it's too late to correlate a quality incident with the actual blend used.

• Every batch starts as a precise blend of wheat lots from different silos: source silo, percentage, moisture after conditioning and the target protein and strength (W) the flour has to hit.
• That sheet lives on paper in the milling control room. It gets transcribed hours later, if at all, by somebody who was not there when the blend was set.
• When the alveograph or the falling number of a batch comes out off-spec, the first question is which silos fed it — and the answer is in a folder, not in a system.
• Without the actual blend tied to the batch, every quality incident is argued from memory, and the head miller's experience never turns into data the next shift can use. Two years later, nobody can say which wheat combination produced the best-performing flour of the season.

How it fits the IRIS system

Connect in photo mode — the miller keeps writing the blend by hand.

Connect photo mode. A grounded LLM extracts structured fields (silos, ratios, target protein/strength, start time) from the sheet photo. The mill supervisor validates with two taps on a tablet before the data crosses into the central system.

Nothing changes in the control room: the sheet is filled in exactly as today and one photo is added. That is why it is usually the first case approved in a mill — it asks nobody to learn a screen. And because every batch from then on carries its blend, the rest of the traceability chain has something to hang from.

See the full IRIS architecture →

Before and after

The blend on paper versus the blend captured

Aspect	Today, on paper	With iLEAN Connect
Silos and percentages per batch	A sheet in the control room	Structured fields in the central system
Capture latency at the flour mill	Hours, or never	Zero lag
Target protein and W	Handwritten next to the ratios	Stored per batch and queryable
Off-spec alveograph result	Which blend? Hunt for paper	One query by batch
Validation	Nobody checks the transcription	Two taps from the mill supervisor
Miller's routine	—	Unchanged: one photo

From hours of latency until manual transcription, to zero lag. From a quality incident with no linked blend, to full lot-blend traceability from minute one.

Impact estimate

Impact estimate — to be validated with your 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.

• Estimated payback 4-9 months, depending on the number of batches blended per day.
• Transcription time recovered by the mill supervisor, who stops retyping sheets at the end of the shift.
• From hours of latency down to zero lag: the blend exists in the system before milling starts.
• And lot-to-blend traceability from minute one, which is what makes any later root-cause analysis possible.

Estimated payback of 4-9 months depending on daily batch volume, recovering the mill supervisor's transcription time. Estimate to validate.

And the fair question from the production manager

"What if it misreads a percentage and the blend in the system is wrong?" — reading fields from a sheet with a known structure is an anchored task, where the best models drop below 1.5% error [1]. And there is a second net: the percentages must add up to 100 and the silos must exist, and the mill supervisor confirms or corrects the summary on the tablet before anything crosses into the central system. A wrong silo number is caught at that moment, not when the batch is already in the bin.

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