Paperless cell start-up: the batch exists at second zero

The problem

The three documents that define the batch live in a folder.

The setup sheet prints with ERP data, but it gets annotated by hand the moment anything drifts: extrusion pressure, crimp height, a blocked cavity, a change of braid reel. The first-piece record is handwritten as a matter of standard, with measured dimensions and the inspector's signature. The shift log captures downtime, reasons and scrap, also by hand. Four to eight hours pass between the signature and the digitization, and considerably more over a weekend. During that gap the batch does not exist for any system: it cannot be queried, cannot be cross-referenced with the leak test, cannot be correlated with anything. The real cost shows up twice. First every morning, when the production meeting argues about how much scrap there was instead of looking at the number. Second on the day a customer opens an investigation and asks for the traceability of one specific batch: then somebody spends hours doing folder archaeology.

• The start-up sheet prints from the ERP, but gets annotated by hand the moment anything deviates: extrusion pressure, crimp height, a blocked cavity, a braid reel change.
• The first-piece record is handwritten by standard, with the measured dimensions and the inspector's signature. The shift report captures stoppages, reasons and scrap, also by hand.
• Between the signature and digitization, four to eight hours pass — and considerably more over a weekend.
• During that gap the batch exists physically but exists in no queryable system, which is exactly when it is needed if a complaint lands.

How it fits the IRIS system

Connect in photo mode — the operator keeps filling in their paperwork.

This is Connect in photo-to-the-analogue mode, the simplest to deploy and the one that creates least friction on the floor: the operator keeps filling in the same paper and only adds two photos with a phone or industrial tablet at start-up. Step 1 — Capture. A photo of the setup sheet and of the first-piece record. No form to fill in, no app to learn. Step 2 — Grounded extraction. A grounded language model extracts structured fields: order number, part number, cavity, process parameters, first-piece dimensions, who signed, downtime with its reason, scrap by cause. Anything unclear in the image is flagged for a human to confirm; nothing is invented. Step 3 — Central memory. The batch is created with its full header and available to be cross-referenced with whatever comes next: leak curve, crimp parameters, certificate for the material consumed, content marked by the laser. Step 4 — Query. Production, quality and management see the same number, by cell and by shift, with nobody having typed it.

It is the simplest of the twelve to deploy and the one that creates least friction on the floor, because it asks for no change: the operator adds a photo to what they already did. That is usually why it is the first one approved.

See the full IRIS architecture →

Before and after

Start-up on paper versus start-up captured

Aspect	Today, on paper	With iLEAN Connect
Capture latency	4-8 h, and more over a weekend	Second zero
Scrap by cause	A figure argued over in the morning meeting	Consolidated data per cell and shift
First-piece dimensions	Handwritten in a folder	History queryable by part number
Deviation noted by hand	Lost unless somebody types it	Extracted and tied to the batch
Answering a complaint	Hunt for paper	One query by batch
Operator's habit	—	Unchanged: one photo

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 start-ups per day.
• Transcription time recovered by production and quality.
• From 4-8 hours of latency — more over a weekend — down to second zero.
• Scrap stops being a figure people argue about and becomes data per cell and shift.

Estimated payback of 4 to 9 months depending on start-ups per day, coming from the transcription time production and quality get back. *Estimate to validate* against the plant's real volumes. Outside that calculation sits an asset that rarely gets valued: the start-up dataset that ends up built, and that feeds every continuous improvement project that follows.

And the fair question from the production manager

«What if it misreads a dimension on the first-piece record?» — reading fields from a document 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 summary goes to the cell tablet and the inspector confirms or corrects it before anything crosses into the system.

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