Paperless Formula Weighing

The problem

The document that proves the formula sits in a folder.

Every formula starts with a paper weighing sheet listing 15 to 30 raw materials, each with its exact quantity and tolerance. It's the document that proves the formula was made correctly, and it can sit untranscribed for hours or get lost in a physical folder.

• A vitamin premix or a fiber-and-protein blend is built from 15 to 30 raw materials, each with a target quantity and a tolerance window. The weighing sheet is where the operator writes the actual weight, the raw material lot and their initials, line by line.
• That sheet is the only proof that the batch was made to formula. Until someone transcribes it, the actual-versus-target deviation of each ingredient exists only in handwriting, on a page that has been next to a scale and a dust extraction hood all shift.
• Transcription happens hours later, when it happens at all: the sheet can sit in a tray until the end of the shift or end up in a weekly binder that nobody ever types up.
• When a customer asks for the composition of a lot shipped three weeks ago, or a carrier lot of maltodextrin is put on hold, the answer means flipping through paper by date and hoping the right sheet is there and legible.

How it fits the IRIS system

Connect in photo mode — the operator keeps the same sheet and adds one photo.

Connect photo mode: the operator photographs the sheet when weighing is done. A grounded language model extracts every ingredient, the actual quantity against target, and the sign-off, with zero change to the routine.

The weighing booth is where the formula becomes real, and it is also where the record is weakest. One photo when weighing closes moves that record from a folder into the central memory, with every ingredient, every actual quantity and every sign-off, and the operator does not have to learn a new screen.

See the full IRIS architecture →

Before and after

The weighing sheet on paper versus the weighing sheet captured

Aspect	Today, on paper	With iLEAN Connect
Time until the sheet is digital	Hours, or never	Zero seconds after the photo
Actual weight against target per ingredient	Handwritten, unreadable at scale	Extracted line by line with its tolerance
Raw material lot on each weighing line	Written if there is room	Tied to the batch and the formula
Finding the sheet of a past batch	Flip through weekly binders	Instant web search by batch or formula
Weighing operator sign-off	Initials on paper	Captured and timestamped
Routine in the weighing booth	—	Unchanged: one photo at the end

Hours of latency until transcription (if it ever happens) → zero seconds. Untraceable sheet → instant web lookup by batch and formula.

Impact estimate

Impact estimate — to be validated with your 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 formulas you weigh per shift and how many weighing booths you run.
• From hours of latency — when the sheet is transcribed at all — down to capture in zero seconds.
• Every weighing sheet becomes searchable by batch and by formula, so a customer composition request is answered with a query instead of an afternoon in the archive.
• Out-of-tolerance lines stop being discovered at the blender or at the lab: the deviation is visible the moment the sheet is photographed, while the batch can still be corrected.

Payback 4-9 months · from hours of latency (if ever transcribed) to zero-second capture — estimate to validate

And the fair question from the production manager

"What if it reads 1.25 kg where the operator wrote 12.5?" — reading fields from a weighing sheet with a fixed layout is an anchored task, where the best models drop below 1.5% error [1]. And there are two more nets: every extracted quantity is checked against that ingredient's target and tolerance, so a misplaced decimal shows up as an impossible deviation, and the summary goes to the booth tablet for the operator to confirm before it is stored.

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