The refinery shift log, digital from second zero

The problem

Two parallel truths live in a modern refinery. Only one of them is queryable.

Two parallel truths coexist in a modern refinery. The control system's — process variables, precise and automatic — and the paper's, which holds the human context: the why. The second is what explains a color deviation three days later, and it is precisely the one that sits in no queryable system. The shift log ends up in a folder in the shift manager's office and does not exist for any system until someone collects it: eight to twelve hours later, at best. When the lab finds that a lot has drifted on moisture, reconstructing what happened means hunting for a piece of paper and calling someone who is already off shift. The result is that root cause analysis, which should be a procedure, is in practice an exercise in collective memory. And as the number of lots and changeovers grows, that memory fails more often.

• The control system's truth — process variables, precise and automatic — and the paper's truth, which holds the human context: the why.
• The second is what explains a colour deviation three days later, and it is precisely the one that sits in no queryable system.
• The shift log ends up in a folder in the shift manager's office and does not exist for any system until somebody collects it — eight hours later, or at the end of the shift.
• By then the answer to "what happened on that strike" is a conversation, not a record.

How it fits the IRIS system

Connect in photo-of-the-analog mode — the operator keeps writing by hand.

Connect in "photo of the analog" mode. Step by step:

Nothing in his routine changes. That is the whole design constraint: a refinery shift manager will not stop mid-round to fill in a form on a screen, and any solution that asks him to is dead by month three.

• The operator photographs his log with a phone or industrial tablet as he closes each block of the shift. Nothing in his routine changes: he keeps writing by hand.
• A grounded language model extracts the structured fields: reading, unit, equipment, time, incident.
• iLEAN links those fields to the active lot and shift, cross-referencing them with what the control system already records.
• The data lands in iLEAN's central memory, queryable by lot, with the photo of the original document linked as evidence.

See the full IRIS architecture →

Before and after

The shift log on paper vs. the shift log captured

Aspect	On paper today	With iLEAN Connect
When the data exists for the system	8 h later, or at shift close	Second zero
The human context — the why	Only on paper	Structured and tied to the lot
Explaining a deviation three days later	A conversation	A query by lot or shift
Cross-reference with the control system	None	Automatic
An illegible or mislaid sheet	The context is lost	The photo remains as evidence
Operator habit	—	Unchanged: still writing by hand

Impact estimate

Impact estimate — to validate against 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 lots and shifts per day.
• From 8 h of latency between the reading and the system, to second zero.
• The shift's human context becomes queryable, which is what every later root-cause analysis rests on.
• Estimate to be validated against the plant's real volumes.

estimated payback of 4 to 9 months depending on the number of lots and shifts per day. *Estimate to be validated* against the plant's real volumes.

And the fair question from the production manager

"What if it misreads a reading and a wrong figure gets in?" — reading fields off a log with a known structure is an anchored task, where the best models drop below 1.5% error [1]. But the real answer is the other one: the extracted summary goes to the tablet and somebody confirms or corrects it before it crosses into the system. Nothing enters on its own.

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