The retort that finally talks

The problem

The most critical asset in the plant still writes its evidence by hand.

the retort's controller typically sits on a network isolated by the manufacturer's own design, and replacing it costs hundreds of thousands of dollars and weeks of downtime on the canning line. Meanwhile, the sterilization curve — the number-one piece of evidence in any food safety audit — gets logged on paper, copied by hand by an operator.

• The retort is the one machine whose curve proves that no spore survived: come-up time, hold temperature, pressure and cooling, cycle by cycle, against the scheduled thermal process.
• Its controller is often fifteen to twenty years old and sits on a network isolated by the manufacturer's own design. Replacing it costs hundreds of thousands of dollars and weeks with the canning line down.
• Meanwhile the cycle log is copied onto paper by the operator reading the panel, with the risk of a transposed digit or a rounded hold time.
• In a food safety audit that log is the number-one piece of evidence, and an incomplete record is a non-conformity even when the load was sterilized perfectly.

How it fits the IRIS system

Connect in photo mode on the retort — from outside the panel, never inside it.

a fixed industrial camera points at the retort's panel; OCR with a grounded language model interprets every frame of the temperature, pressure and time curve, cross-checking it with whichever batch the ERP shows as active in the retort at that moment.

The camera reads what the operator reads, only every frame and without transcribing. The retort keeps its controller, its validation and its manufacturer support; what changes is that each load leaves a complete digital curve tied to the batch the ERP had in the retort at that minute.

See the full IRIS architecture →

Before and after

The hand-copied cycle log versus the camera-read curve

Aspect	Today	With iLEAN Connect
Cycle record	Copied onto paper by the operator	Full curve, frame by frame
Come-up and hold phases	A few readings noted	Every point of temperature, pressure and time
Which batch was inside	Written by hand, if at all	Cross-checked with the ERP at that minute
A deviation from the scheduled process	Spotted when the log is reviewed	Flagged while the cycle runs
The retort's controller	Isolated and untouched	Still isolated and untouched
New retort capex	Pressing	Deferred

hand-copied paper log with error risk → a complete, traceable digital curve per batch with no human intervention.

Impact estimate

Impact estimate — to be validated with your retort records.

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, combining avoided capex on a new retort with lower non-conformity risk from incomplete records.
• The controller replacement, hundreds of thousands of dollars plus weeks of canning line downtime, stops being the only way to get a digital record.
• Non-conformities from incomplete or illegible cycle logs drop, because the record no longer depends on someone copying it by hand.
• And any question about a load is answered with its complete curve, not with a sheet that says the cycle was fine.

estimated payback 5-10 months, combining avoided CAPEX on a new retort with reduced non-conformity risk from incomplete records. *Estimate to validate.*

And the fair question from the production manager

"Can a camera pointed at a screen be trusted with a sterilization record?" — reading a fixed panel, with values always in the same position and a physically bounded range of temperature and pressure, is an anchored task where the best models stay under 1.5% error [1]. A reading outside the possible range is rejected, not stored: the system flags the frame instead of writing a temperature the retort could never have reached. And the paper log can run in parallel during the first weeks, until quality trusts the curve.

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