The blow molder start-up recipe, captured in one photo

The problem

The most valuable document of the shift is the worst kept.

The set-up sheet is the most valuable document a PET plant produces in a shift, and the worst preserved. It lives in a binder next to the blow molder or a folder in the production office, handwritten, with crossings-out and margin notes only its author fully understands. The cost does not show up in the accounts, it shows up on the clock. Every time a reference comes round again, the technician gropes for the setting once more: oven profile, pressures, delay. Meanwhile everything coming off the machine goes to regrind. You pay twice — and three and four times — for the same learning. There is a second, slower and more expensive cost: the knowledge is never capitalised. In a group with several plants and new openings, what is not written down cannot be replicated. A new facility starts its learning curve from zero because the mature plant's experience never left the paper.

• The setup sheet ends up being the real recipe of the batch: not the theoretical one in the system, but the one that worked that day with that resin and that recycled content.
• It lives in a folder next to the blow molder or in a drawer in the production office, handwritten, with crossings-out and margin notes only its author understands.
• The cost does not show up in the accounts, it shows up on the clock: every time the part number comes back, the technician gropes for the setting again — oven profile, pressures, stretch delay.
• And meanwhile everything coming off the machine goes to regrind. It is paid twice, and three times, and four.

How it fits the IRIS system

Connect in photo mode — the technician keeps writing on paper.

Connect, "photo of the analogue" mode. The technician photographs his sheet with a phone or industrial tablet at start-up. Nothing about his routine changes: he keeps writing on paper. Step by step:

The recipe ends up tied to the mold, the resin lot and that day's recycled content. That tie is what turns a sheet of paper into a starting point: next time the part number runs with that blend, the start-up does not begin from zero.

• Capture. A photo of the set-up sheet and, optionally, of the first preform and first bottle of the batch.
• Extraction. A grounded language model recognises the document structure and extracts the fields: mold, active cavities, oven profile, pre-blow and blow pressures, stretch rod delay, target weight, resin batch, recycled content and handwritten notes in the margin.
• Early human verification. The summary appears on the line tablet. The technician confirms with two taps or corrects before the data goes in. Nothing crosses into the central system without a human signature.
• Central memory. The recipe is tied to the mold, the resin batch and the recycled content of that day, and cross-referenced with drying conditions and the lab tests for the same batch.
• Reuse. Next time that format runs, the system proposes the recipe that already worked with an equivalent resin.

See the full IRIS architecture →

Before and after

Start-up on paper versus start-up captured

Aspect	Today, on paper	With iLEAN Connect
Data latency	4-8 h, or never digitized	Second zero
Start-up starting point	Groping from zero	The recipe that already worked
Where the knowledge lives	With whoever is on shift	In a queryable history
Tied to resin and recycled content	No	Yes, mold by mold and lot by lot
Kilos in the start-up tail	To regrind, every time	Shorter with every repetition
Technician's routine	—	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 changeovers per month.
• A shorter start-up tail: kilos of resin that stop going to regrind.
• Less dependence on the specific person who knows how to find the setting.
• From 4-8 hours of latency — or never being digitized — down to second zero.

Estimated payback 4-9 months, depending on how many changeovers run per month. The return arrives two ways: a shorter start-up queue (kilos that stop going to regrind) and reduced dependence on whoever happens to be on shift. *Estimate to be validated* against the plant's real changeover data.

And the fair question from the production manager

«What if it misreads a pressure or an oven profile?» — 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 appears on the line tablet and the technician confirms with two taps or corrects before the data enters. Nothing crosses unsigned.

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