The autoclave panel, digitized without touching the machine

The problem

The link where most of the yield is decided leaves no trace.

The autoclave is the bottleneck and the first link in the yield chain. Two nominally identical loads can behave differently depending on agave maturity, how it was loaded and how pressure evolved. Nobody can prove that today, because the curve is not stored: at best someone writes "normal cook" by hand. The classic answer would be to replace the equipment or ask the manufacturer to open up its controller. That is six figures of investment and weeks of plant downtime, on a machine that works perfectly well. Nobody does that just to get data, and rightly so. The result is a blind spot at exactly the link where most of the agave yield is decided.

• Cooking the piñas decides how much inulin converts to fermentable sugar and which aromas appear. The autoclave is the bottleneck and the first link in the yield chain.
• Two nominally identical loads behave differently depending on agave maturity, how the piñas were loaded and how pressure evolved. Nobody can prove that today, because the curve is not stored: at best somebody writes "normal cook" by hand.
• The manufacturer's panel shows the curve while the cycle runs and forgets it when the cycle ends. Somebody may glance at it and remember; the plant does not.
• The classic answer — replacing the equipment or asking the manufacturer to open its controller — is six figures and weeks of downtime on a machine that works perfectly well. Nobody does that just to get data, and rightly so.

How it fits the IRIS system

Connect in photo mode on the panel — solved from outside, not inside.

Connect in photo-on-panel mode. The flow:

The camera is not connected to the autoclave: it looks at it. No PLC change, no software change, no warranty conversation with the manufacturer. It is exactly the same autoclave; it just speaks now, load by load, and the master distiller reads the curve instead of remembering it. He does not run the cook differently; he runs it with a memory. And nobody has to stand at the panel to write anything down.

• A fixed industrial camera is installed pointing at the autoclave panel. It is not connected to the machine: it looks at it.
• It captures every few seconds.
• OCR with a grounded LLM interprets each reading — pressure, temperature, time, cycle phase, alarms — and reconstructs the full time curve of the load.
• The curve is automatically tied to the active batch (case 1) and cross-referenced with must brix and lab analysis (case 6).

The manufacturer's PLC is untouched, its software is untouched and the warranty stays intact. It is exactly the same autoclave: it just speaks now.

See the full IRIS architecture →

Before and after

Today's autoclave versus the autoclave being read

Aspect	Today	With iLEAN Connect
The cooking curve	Gone when the cycle ends	Digitized per load
Comparing two loads	"Normal cook" in two handwritings	Curve against curve, on one chart
Correlation with brix and liters	Impossible	Automatic, tied to the batch
The manufacturer's PLC	—	Untouched, warranty intact
Replacement CAPEX	In the plan	Deferrable indefinitely
Alarms and phase changes	Seen by whoever passed by	Timestamped in the load history

From no cooking history at all to a full curve per load, comparable across loads and correlatable with brix, liters and sensory profile. Autoclave replacement deferrable indefinitely.

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 5-10 months.
• Avoided CAPEX of replacing or upgrading the cooking equipment, which stays exactly as it is, running exactly as the master distiller runs it today.
• Reduction in yield dispersion between loads, once you can see which curves yield better and repeat them on the next load of agave from the same field.
• And the cooking history the plant never had becomes the first column of every yield conversation.

Estimated payback 5-10 months, counting the avoided CAPEX of replacing or upgrading the cooking equipment and the reduction in yield dispersion between loads once you can see which curves yield better. *Estimate to be validated.*

And the fair question from the production manager

"Can a camera really read a panel reliably enough to trust the curve?" — here the task is anchored: a fixed screen, fields in known positions and physical ranges the value cannot leave, where the best models drop below 1.5% error [1]. And any frame whose reading falls outside the possible range for pressure or temperature is discarded instead of accepted, so an artifact never enters the series. If the panel is partly obscured by glare or steam, the missing seconds are flagged as a gap, not filled in.

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