The master distiller talks while he walks

The problem

The finest information in the plant dies at the end of the shift.

In a premium distillery the critical knowledge is sensory and contextual. No sensor holds it: it depends on a person who has spent decades smelling, tasting and looking. That tank six started slow, that the coil is running tight on temperature, that a barrel in row three is sweating more than usual. That person cannot type: their hands are full and they are moving. And asking them to fill in a form at the end of the shift is asking them to reconstruct from memory what they noticed eight hours earlier. What gets lost is precisely the observations that anticipate a problem — the ones that arrive as instinct rather than as an alarm. What survives today are stray comments in a chat group, unstructured, tied to no tank and no batch, impossible to look up when the same pattern repeats three months later.

• In a premium distillery the critical knowledge is sensory and contextual. No sensor holds it: that tank six started slow, that the coil on the copper still is running tight on temperature, that a barrel in row three is sweating more than usual.
• The person who holds it cannot type: he has a hydrometer in one hand and a flashlight in the other and he is moving between cooking, tanks, stills and the barrel cellar.
• Asking him to fill in a form at the end of the shift is asking him to reconstruct from memory what he noticed eight hours earlier. What gets lost is precisely the observation that anticipates a problem.
• What survives today are stray comments in a chat group, tied to no tank and no batch, impossible to look up when the same pattern repeats three months later.

How it fits the IRIS system

Connect in voice mode — iLEAN listens to the master distiller, not to the plant.

Connect in voice mode. The flow:

The microphone activates when he wants it to, and only for what he chooses to say. That distinction is not cosmetic: it is what keeps the case alive in month six instead of being perceived as monitoring. And the answer that comes back in seconds — that tank also started slow in the two previous batches with the same yeast lot — is support, not a decision. He can also ask — which tanks started slow this year? — and get the list, with the batches beside it.

• A light Bluetooth earpiece, of the kind already used on the floor.
• The master distiller activates the microphone when he wants and dictates in plain language: "note that tank six has been running two days slow since the yeast lot changed".
• iLEAN structures the observation, ties it to the active tank and batch, and stores it in central memory.
• And it answers in seconds with what it already knows: "that tank also started slow in the two previous batches with that same yeast lot — shall I flag it to the lab?"

One design point matters here: the AI listens to the conversation the master distiller directs at it, not to the plant and not to the operators. The microphone activates when he wants it to. Humans in command, without exceptions.

See the full IRIS architecture →

Before and after

Today's round versus the captured round

Aspect	Today	With iLEAN Connect
A qualitative observation	Lost at the end of the shift	Tied to tank, batch and yeast lot
The master distiller's hands	Full: hydrometer and flashlight	Still full: he just talks
A slow tank	"I think this happened before"	"It happened in these three batches"
Where the notes live	A chat group	Searchable central memory
What comes back	Nothing	What the house already knows, in seconds
Operators' conversations	—	Not recorded

From no record of qualitative observation at all to a history searchable by tank, by batch and by yeast lot. From "I think this happened before" to "it happened in these three batches, with this common factor".

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, hard to isolate on its own because its main value is preserving the master distiller's knowledge as an asset of the house.
• A history searchable by tank, by batch and by yeast lot where today there is no record at all.
• Problems anticipated from an observation instead of discovered from an alarm — a slow tank flagged on day two rather than at the still.
• And the sensory criteria of the person who decides the cut, available to the house the day that person is no longer there.

Estimated payback 5-10 months, hard to isolate because its main value is preserving the master distiller's knowledge as an asset of the house. *Estimate to be validated.*

And the fair question from the production manager

"Will it understand him with the pumps running and trade words nobody outside the plant uses?" — recognition is prepared for industrial noise and for the vocabulary of the house: piña, ordinario, heads, tails, the names of the tanks. Structuring a dictated observation into tank, batch, symptom and severity is an anchored task, where the best models drop below 1.5% error [1], and what was understood is shown back to him so he can correct it with one word.

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