PAPERLESS RECIPE STARTUP

The problem

The sheet that fixes salt, fat and starch stays on paper all shift.

Every batch startup in a bouillon and soup factory begins with a paper production order and a recipe sheet fixing salt, aromas, starch and fat for that specific format and brand. With several brands sharing the same dosing unit across shifts, both documents can go 4 to 8 hours without being digitized. If a flavor or texture complaint comes in, reconstructing the exact recipe used for that batch depends on finding the right paper in the dosing supervisor's folder — a slow process exactly when speed matters most.

• The production order and the recipe sheet fix the proportions of salt, aromas, starch and fat for one brand and one format, and both travel to the dosing unit on paper.
• With several brands sharing the same dosing unit across shifts, the two documents can go 4 to 8 hours without being digitized, and over a weekend considerably longer.
• During that window the batch has been blended, pressed, filled and cartoned, but it exists in no system anybody can query, and the run is already on a pallet.
• When a flavor or texture complaint lands, reconstructing the exact recipe means finding the right sheet in the dosing supervisor's folder — slow precisely when speed is what you need, and slower still if the shift that ran it is off that week.

How it fits the IRIS system

Connect in photo mode — the operator adds a photo to what they already do.

Connect photo mode: the operator takes two phone photos of the order and the recipe sheet at startup. A grounded LLM extracts structured fields (batch, brand, format, recipe proportions, declared allergens) without changing the operator's routine. The data passes through the early human-verification tablet before entering the central system.

Nothing changes in the dosing routine: the same paper, the same signature, one extra photo before the blender starts. That is why this is usually the first of the twelve a plant approves, the fastest to see working, and the one that makes the other eleven credible to the people who have to run them.

See the full IRIS architecture →

Before and after

The recipe sheet today versus the recipe sheet captured

Aspect	Today	With iLEAN Connect
Time to digitize the recipe	4-8 hours	Zero seconds
Salt, fat and starch proportions	Only on the paper sheet	Structured fields on the batch
Brand and format of the run	Inferred from the schedule	Read off the order itself
Allergens the recipe declares	Written on the sheet	Tied to the batch from the start
A flavor or texture complaint	Search the dosing folder	The exact recipe in one lookup
Dosing supervisor's transcription	Every batch	None

4-8 hours of latency before the batch recipe is digitized → the batch exists in the central system in zero seconds, with its exact recipe instantly searchable.

Impact estimate

Impact estimate — to be validated with your own batch figures.

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 your dosing unit starts per day.
• The dosing supervisor's transcription time comes back whole: it is the hour that never appears in the shift's cost because nobody bills it to anything.
• From 4-8 hours of latency to a batch that exists, with its exact recipe and its declared allergens, from the moment it starts.
• And answering a complaint stops being an archaeology exercise in a folder: the recipe used is a lookup by batch number, whoever is on shift that day.

Estimated payback of 4 to 9 months depending on daily batch volume, fully recovering the dosing supervisor's transcription time. Estimate to be validated.

And the fair question from the production manager

"What if it reads a proportion wrong and we dose badly?" — extracting fields from a sheet with a fixed structure is an anchored task, and on anchored tasks the best models drop below 1.5% error [1]. The second net is that the extraction never touches the doser and never changes a setpoint: it goes to the line-side tablet as a summary, and the dosing supervisor confirms or corrects it before it crosses into the central system. The recipe that governs the blend is still the paper one in their hand.

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