Multi-product golden batch: every SKU has its perfect batch — and today it lives in the veteran operator's head.

The problem

The recipe is written down. The good batch is not.

Every process plant has an approved procedure per SKU: temperature ranges, minimum and maximum times per stage, dosing tolerances. That document defines the permitted window. What it does not define is the optimal path within that window — and that is where the margin is won or lost.

Two batches of the same SKU can comply with the entire procedure and come out very different: one with yield at the top and no rework, the other right on the edge of the specification. The difference lies in details the procedure does not capture: how fast the temperature was raised, how long the wait was before dosing, how much margin the stage was closed with. Details a control-room operator with years of experience does by instinct and, when asked, sums up as “you can tell when it's about ready.”

Out of that come four problems the operation pays for every week:

1. The knowledge is not transferable — the morning shift turns out one product and the night shift a slightly different one, with the same recipe and the same machine. It is not a lack of rigor: the good pattern is simply not written anywhere.
2. The product changeover is a blind spot — the cleaning, the purge and the stabilization until the new SKU comes out in spec generate scrap that is almost never charged to anyone, and that varies enormously depending on who is in the control room.
3. There is nothing to compare against in real time — the control system warns if you leave the hard limits, but not if you are straying from the path that produced the best batches. By the time the hard limit trips, the correction is already expensive.
4. The analysis arrives after the fact — the out-of-spec batch is investigated days later, with data that has to be reconstructed from three different systems, and the conclusion is usually “it drifted somewhere in stage 3.” Without the good batch's curve alongside, that is an autopsy, not a correction.

A multi-product plant multiplies all of this by the number of SKUs. With fourteen products passing through the same reactor there is no single golden batch: there are fourteen, each with its own curve signature. The goal is not to write the procedure better. It is to make the batch that already came out right explicit, measurable and comparable.

How it fits the IRIS system

iLEAN does not invent the good recipe — it extracts it from your own batches and puts it in front of the operator.

The golden batch is, at heart, standard work applied to a continuous process: find the best known method, make it visible and use it as the base for the next improvement. What kept it from being applied in multi-product plants until now was purely practical — no one has the time to reconstruct by hand the curves of the best batches for fourteen SKUs. That is exactly what an agent does without tiring. Connect is the putty that fills the cracks between the historian, the SCADA, the production order and the lab result.

The new operator stops wondering whether things are on track. They have the best batch's curve for that SKU in front of them and can see, stage by stage, whether they are above or below it. The veteran's experience stops being a fragile asset.

The iLEAN pieces applied to the multi-product golden batch:

• Edge — captures in-line the variables that are not instrumented or whose data never reaches the history: product appearance by vision, readings from an old panel, the stage's real state. Raw data in milliseconds. It works without a network.
• Connect — joins the historian, the SCADA, the MES production order, the lab result and the shift lead's spreadsheet into a single timeline per batch. Without this unification there is no golden batch possible: the curves exist, but they do not line up with each other.
• Agents — the brain. They score each SKU's historical batches on your criteria, isolate the group of the best ones, extract the reference profile with its dispersion band stage by stage, load the correct profile when each batch starts and compare continuously. When the separation stops being noise and becomes a trend, they alert.
• Central memory — every finished batch enters the history and readjusts the profile. If one shift finds a better way to run a SKU, that way becomes the new profile for every shift, not a hallway anecdote.
• Three safety rings — no proposal acts on the process. The profile is approved by the process lead; the correction is signed by the operator. The system proposes, the person decides.

The golden batch rests on the same layers as two other solutions in this family, and it pays to look at them together: real-time SPC with AI provides live statistical control over the batch's critical variables — charts, capability, drift rules — and predictive quality with AI turns the detected deviation into a prediction of the batch's final result before it finishes. The golden batch answers “am I on this SKU's good path?”; SPC answers “is the variable under control?”; predictive quality answers “how is this going to come out?”. They are deployed separately and reinforce each other when they coexist.

See the full IRIS architecture →

Before and after

Written procedure + the veteran's judgment vs. a per-SKU golden batch with iLEAN

Aspect	Today: procedure + experience	With the iLEAN golden batch
Comparison reference	The procedure's window (hard limits)	The profile of that SKU's best real batches
Multi-product	One general criterion for the whole line	One profile per SKU, loaded automatically when the batch starts
Product changeover	Depends on who is in the control room; variable scrap	Reference sequence for cleaning, purge and stabilization with an expected duration
Drift detection	When the hard limit trips or someone notices	As soon as the curve leaves the profile's band and the separation is a trend
Alert	Verbal, inside the control room	To the operator's terminal and the shift lead's channel, with the chart and the stage
Analysis of the out-of-spec batch	Days later, reconstructing data from three systems	Immediate, with the reference batch's curve overlaid
The veteran's knowledge	Leaves with the person	Stays in the central memory and is inherited across shifts
Profile improvement	Occasional, project by project	Continuous: every good batch readjusts the profile

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Process plant with several SKUs on the same line or in the same reactor, three shifts and recurring product changeovers. History available, even if scattered across systems.
• Edge + Agents pilot on one line: profile extraction for the highest-volume SKUs, real-time comparison and alerts to the terminal. First usable profile within a few weeks if history exists; first value on the very next startup.
• Indicative payback between 5 and 10 months, from two main levers: fewer out-of-spec batches (less rework, fewer downgrades to second quality) and less scrap on product changeovers (shorter, more repeatable startups).
• Secondary levers that usually surface in the diagnostic: reduced shift-to-shift variability, lower energy and auxiliary-material consumption in the long stages, and a noticeably shorter learning curve for the new operator.

And the process lead's reasonable doubt

“What if the system flags a drift that does not exist and I end up over-correcting?” — it is the right objection, and it has two answers. The first is about method: comparing a curve against a historical profile with its dispersion band is an anchored task, not a generative one. There is a data series, a reference profile and a numerical separation criterion; there is no room for the system to make anything up. Hallucination is a problem of free generation, and on anchored tasks the best models are below 1.5% error ^[1]. The second answer is about design: the alert reaches the operator as a proposal, with the chart and with what the good batches did at that same point. No one corrects blindly because a screen turns orange — they correct because they see the separation and understand it. And if the operator believes the profile is miscalibrated for that SKU, that disagreement also enters the system and refines the profile.

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

You may also be interested in: Real-time SPC with AI · Predictive quality with AI · DMAIC Six Sigma with AI statistical analysis