Batch start-up stops living on a sheet of paper

The problem

The most valuable part of the document is the part written by hand.

The manufacturing order is printed from the ERP and annotated by hand as soon as there is a real deviation: a substituted raw-material lot, an adjusted quantity, a supervisor's remark. That handwriting is the most valuable part of the document, because it is where what actually happened gets recorded — and it is exactly the part that lives in no system. Four to eight hours, or a whole shift, can pass before anybody digitises those sheets. Meanwhile the batch exists physically but does not exist in the data. And when months later somebody has to investigate why a shade came out off standard, the reconstruction starts by hunting for paper in a filing cabinet.

• The production order prints from the ERP, but what actually happened is written on top of it by hand: the substituted raw material lot, the adjusted quantity, the supervisor's note.
• That handwritten note is the first evidence a GMP auditor asks for, and it is exactly the part that lives in no system.
• For four to eight hours, or the whole shift, the batch exists physically but does not exist in the data.
• Months later, when someone has to investigate why a shade came out off-standard, the reconstruction starts by hunting for paper in a filing cabinet.

How it fits the IRIS system

Connect in photo mode — the technician keeps filling in the same paperwork.

Connect in photo-of-the-analogue mode. The technician photographs the order and the weighing sheet with a phone or the weighing-room tablet. A grounded LLM recognises the structure of the document and extracts the fields: shade, formula code, quantity, lot of each raw material, actual weight, deviation from theoretical, technician signature and handwritten remarks. The result goes through the early human verification tablet and, once signed, is stored in central memory tied to the batch identifier. Flow: photo → structured extraction → visual summary on the tablet → technician ✓ → central memory → retrievable by batch number, with the image of the original sheet attached as evidence.

The habit change is zero, and in a GMP weighing room that is not about convenience: it is the only way the record still exists in month six. Anything that asks for typing between weighings gets filled in at the end of the shift, from memory.

See the full IRIS architecture →

Before and after

Start-up on paper versus start-up captured

Aspect	Today, on paper	With iLEAN Connect
When the batch exists in data	4-8 h later, or at shift close	Second zero
The handwritten note	Lost, or filed unread	Extracted and tied to the batch
Raw material → product traceability	Rebuilt by hand	A graph that builds itself
GMP start-up evidence	Available weeks later	Available the same day
Investigating an off-standard shade	Starts by hunting for paper	Starts with the data in front of you
The weighing technician's routine	—	Unchanged: two photos

Impact estimate

Estimated impact — to validate with your own 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 per day.
• Transcription and filing time recovered by the production supervisor.
• From 4-8 h of latency between the paper and the system, down to second zero.
• Start-up evidence available the same day instead of weeks later.

estimated payback of 4 to 9 months depending on batches per day, counting the transcription and filing time recovered by the manufacturing supervisor. *Estimate to validate*.

And the fair question from the production manager

«What if it misreads a raw material lot?» — reading fields from a document with a known structure is an anchored task, where the best models drop below 1.5% error [1]. But the real answer is the other one: the extracted summary goes to the tablet and a person confirms or corrects it before it crosses into the system. In a GMP environment that is not optional, and here it is the architecture.

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