The batch exists in the system from minute one

The problem

The batch's rating lives on two sheets of paper nobody has typed yet.

At a kraft fiber drum plant, the production sheet is filled in by hand with the order's data — format, closure type, required dangerous-goods rating — and the first-drum control label is also handwritten by quality policy. Both documents can sit undigitized for hours, piling up in the production lead's office folder. If someone needs to reconstruct which rating a given batch carried, today that means tracking down the physical paper — and sometimes it isn't where it should be.

• The production sheet is filled in by hand with the order's data: drum diameter and height, closure type, steel hoop or not, liner, and the dangerous-goods rating the order requires.
• Quality policy also asks for the first-drum control label to be handwritten, with batch, line, time and the operator's initials. It is the reference for the whole batch if anything goes wrong later.
• Both documents can sit undigitized for hours, piling up in a folder in the production lead's office while the batch is already being wound, assembled and palletized. Nobody outside that office can see what is being made or for which rating.
• If someone has to reconstruct which rating a given batch carried, today that means tracking down the physical paper — and sometimes it isn't where it should be.

How it fits the IRIS system

Connect in photo mode — two photos at startup, and the paper stays as it is.

With iLEAN Connect's photo mode, the operator takes two phone photos at batch startup: the production sheet and the control label. An anchored LLM recognizes the document structure and extracts batch, format, closure type, required rating and notes, storing them in central memory at zero latency — without the operator changing anything in their routine.

The operator keeps filling in the sheet and the label exactly as quality requires. The only new gesture is two photos, and from that moment the batch, its format and its rating can be queried from anywhere in the plant instead of from a folder on a desk. When a chemical customer calls about a batch, the answer is on a screen in seconds, with the photo of the original sheet attached.

See the full IRIS architecture →

Before and after

Drum batch startup on paper versus startup captured by photo

Aspect	Today, on paper	With iLEAN Connect
Time until the batch exists in the system	Hours, sometimes a full shift	Minute one
Required UN rating per batch	Written on a sheet in a folder	Extracted and tied to the batch
Closure type and drum format	Readable only on paper	Structured fields, searchable
First-drum control label	Handwritten and filed	Photographed, read and linked
A sheet that goes missing	The rating can't be proved	The photo is already stored
Operator's routine	Fill in and file	Fill in, two photos, carry on

Before: hours of latency before the sheet gets digitized, paper that sometimes goes missing. After: the batch is queryable from minute one, with its rating traced by photo.

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 how many batches you start per day.
• It comes from reclaiming the production lead's transcription time, which today goes into typing sheets that have already been written once.
• From hours of latency to a batch that can be queried from minute one, with its rating traced by photo.
• And the question "which rating did that batch carry?" stops depending on finding a sheet of paper.

Estimated payback 4-9 months depending on batch volume per day, from reclaiming the production lead's transcription time. Estimate to be validated.

And the fair question from the production manager

“What if it misreads the packing group on a handwritten sheet?” — reading fields from a document whose structure is known is an anchored task, where the best models drop below 1.5% error [1]. And there is a second net: the rating is checked against the active order in the ERP, and any mismatch goes to the line tablet for a person to confirm before it is stored. A rating that disagrees with the order is never filed silently.

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