Paperless line startup

The problem

The two sheets that decide the batch sit in a folder until lunch.

The Production Order prints each morning from the ERP, and any last-minute adjustment gets scribbled on by hand. The oven startup log is handwritten by habit. Both can sit undigitized for four to eight hours — if there's an oven deviation mid-morning, nobody correlates it with the startup until the next day, and root-cause analysis arrives too late to prevent it recurring in the next batch.

• The Production Order prints from the ERP every morning, but the real batch is defined by what gets scribbled on it: a target weight nudged up, a recipe swapped, a belt speed changed because the mixer ran late.
• The oven startup log is handwritten by habit: the temperature of each tunnel oven zone, belt speed, the time the first pans went in and the line leader's initials.
• Both sit undigitized for four to eight hours. If the tunnel oven drifts mid-morning, nobody ties it to how it was started until the next day, when the folder finally reaches the office.
• By then the next batch has already been baked with the same settings, and the root-cause analysis arrives one batch too late to prevent the repeat.

How it fits the IRIS system

Connect in photo-of-paper mode — the line leader keeps the clipboard and adds a photo.

Connect's photo-of-paper mode. The line leader photographs the order and the startup log with their phone. A grounded language model extracts format, target weight, recipe, initial zone temperature and belt speed, storing it in iLEAN's central memory with nobody transcribing anything. Zero habit change for the line leader.

Nothing on the line changes: the same order, the same log, the same pen. What changes is that the batch exists in central memory from the first minute, with the handwritten adjustments included, which is exactly where the real story of the shift is written, and where the oven deviation of mid-morning finds its explanation. That is usually why it is the first case a bakery approves.

See the full IRIS architecture →

Before and after

Startup on the clipboard versus startup in central memory

Aspect	Today, on paper	With iLEAN Connect
When the batch is queryable	4-8 h later, when someone collects the folder	Zero latency, from the two photos
Target weight changed by hand	Lives only on the printed sheet	Extracted and tied to the batch
Initial oven zone temperatures	Handwritten, never charted	Stored per batch and per oven
Oven drift at mid-morning	Correlated the next day, if ever	Visible in the same shift
Belt speed and recipe at startup	Inferred from memory	Recorded exactly as started
The line leader's routine	—	Unchanged: two photos

Impact estimate

Impact estimate — to be validated with your bakery's 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 line startups you run per day.
• Transcription time recovered for the line leader, who today copies the same figures into the system at the end of the shift or never does.
• From four to eight hours of latency down to zero: an oven deviation becomes visible in the same shift instead of the next morning.
• Root-cause work on a bad batch starts from the real startup settings, not from what someone remembers about them.

Estimated payback of 4 to 9 months depending on daily startup volume, plus recovered transcription time for the line leader. Estimate to validate against real plant data.

And the fair question from the production manager

“What if it misreads a zone temperature the line leader wrote by hand?” — reading fields from a document with a known structure is an anchored task, where the best models drop below 1.5% error [1]. On top of that, every temperature and speed is checked against the physical range of that oven, and the summary reaches the line tablet for a person to confirm before anything is stored as final.

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