The harvest log, digitized at zero latency

The problem

The document that decides the destination tank sleeps in a folder.

During harvest peak, dozens of loads arrive every day, each with its own handwritten log. The volume overwhelms manual transcription in real time — the tank's data stays blind for hours, right when decisions need to be fastest.

• Next to the scale, the cellar master writes the intake log by hand for every trailer: variety, plot, grower, gross weight, probable alcohol reading and the tank the grapes are sent to.
• That sheet decides where each load ends up, but it only becomes data when someone types it in at the end of the day — or the next morning, after a twelve-hour shift at the weighbridge.
• At harvest peak, dozens of loads arrive every day, each with its own handwritten log. Transcription cannot keep up, and the tank fills physically while it is still empty in the system.
• That blind window lands exactly when decisions need to be fastest: which tank still has room, which plot is coming in under-ripe, which variety needs a separate fermentation, which load should go to the reserva tanks.

How it fits the IRIS system

Connect in photo mode — the intake log stays on paper, and it is in the system too.

Connect photo mode. The team photographs the log with a phone or tablet. A grounded LLM extracts structured fields (variety, plot, origin, weight, destination tank, timestamp) and stores them in central memory. Zero habit change during the tightest week of the year.

Nobody is asked to change anything during the tightest week of the year: the paper log stays, the photo is added. That is why this case is usually the first one a winery approves, and why it becomes the foundation every later case reads from — the batch starts existing at the weighbridge.

See the full IRIS architecture →

Before and after

The intake log on paper versus the intake log captured

Aspect	Today, at the scale	With iLEAN Connect
When the load exists in the system	End of the day, or the next morning	Zero latency, at the weighbridge
Tank occupancy during peak	Known by whoever is on the floor	Visible to the whole cellar team
Probable alcohol per plot	Handwritten, hard to compare	Comparable across plots and campaigns
Variety assigned to each tank	Depends on reading the handwriting	Structured field, tied to the batch
Transcription work	Hours per day during harvest	None
The cellar team's routine	—	Unchanged: one photo per log

hours of lag during peak campaign → zero latency; tank with no digital record until the next day → traceability from the weighbridge itself.

Impact estimate

Impact estimate per campaign — 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 the number of grape loads per campaign.
• The return comes from transcription time recovered for the cellar team, right in the weeks when that team is most stretched.
• From hours of lag during peak campaign to zero latency between the scale and the central system.
• And traceability starts at the weighbridge itself, not the next day when the tank already holds the grapes and nobody remembers which trailer came first.

estimated payback of 4-9 months depending on load volume per campaign, from recovered transcription time for the cellar team. Estimate to validate.

And the fair question from the production manager

"What if it misreads the variety or the weight on a smudged log?" — reading fields from a sheet whose layout is known is an anchored task, where the best models drop below 1.5% error [1]. And the second net is physical: a weight outside what the scale can register, or a variety the plot has never grown, is flagged for a person to confirm before anything is stored.

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