Flash juice pasteurization with AI — the HTST window is seconds wide, and the decision cannot take longer.

The problem

The time-temperature pair is the backbone — and it is managed with two screens and a hunch.

In a flash juice plant, the quality technician cross-references three realities in their head every shift:

1. The nominal HTST curve — the one the SKU's HACCP plan dictates (for example, 92 °C × 15 s for orange juice with pH < 4.5).
2. The real curve — the one the pasteurizer's SCADA draws, which fouls up with pulp, changes with pump flow and drifts as the heat exchanger ages.
3. Sensory quality — color, aroma, residual vitamin C, which is only confirmed in the lab hours later, when the batch is already packed.

The classic system works because there is a procedure. But nobody knows the effective time-temperature pair of batch 412 at 3:23 p.m. with any precision until somebody patiently reconstructs it, and oxidation caused by residual O₂ in an old valve shows up in the color when you are already in distribution. iLEAN does not invent the control; it closes it at a speed one person alone cannot sustain shift after shift.

How it fits the IRIS system

iLEAN does not replace your pasteurizer or your SCADA — it seals the cracks between the nominal and the real.

The problem in flash juice is not a lack of instrumentation: it is that the real curve, the flow, the pH and the color each sit on a different screen, and they are only cross-referenced at the end, when the batch is already closed. iLEAN acts as the putty that joins those islands and closes the see-decide-act loop before packing.

Edge sees the color at the sight glass. Connect reads the curve, the flow and the recipe. The agent cross-references them with the HACCP time-temperature pair and, if something does not add up, proposes a hold. The person signs — never the other way round.

The three iLEAN pieces applied to flash juice pasteurization:

• Edge — a terminal with vision over the pasteurizer's outlet sight glass and over the packing line (label, date, batch). It detects the color/saturation delta that indicates over-cooking or oxidation, and verifies the printed label. It works with no network: if the plant loses WiFi, Edge keeps classifying and holding.
• Connect — captures the HTST curve from the SCADA, the pump flow, the inlet pH, the batch recipe and the record of the previous CIP, whether they come from a modern SCADA, an old PLC or the process manager's spreadsheet. And it captures what arrives from outside: the supplier's email about a new orange variety, the maintenance manager's WhatsApp about a heat exchanger swap.
• Agent — calculates the effective time-temperature pair batch by batch, compares it with the SKU's HACCP curve and proposes a hold if the pair falls below the CCP. For routine audits, it prepares the complete per-batch dossier (real curve, flow, pH, CIP, color, signature). The Agents have no hands on the critical work order: they propose and the responsible person signs.

See the full IRIS architecture →

Before and after

Nominal curve + lab afterwards vs. effective time-temperature pair cross-referenced with iLEAN

Aspect	Nominal curve + lab afterwards	With iLEAN Edge + Connect + Agent
Effective pair for the batch	Assumed = nominal	Calculated batch by batch, compared with the CCP
Detecting oxidation / over-cooking	In the lab, hours later	At the sight glass, in line, before packing
Heat exchanger fouling	Scheduled by calendar	Anticipated by the drift in the real curve
Verifying the intermediate CIP	Checklist signed per shift	Cross-referenced with the SKU changeover and traced by batch
Energy consumed per SKU	Conservative setpoint at all times	Optimal setpoint per SKU without losing the CCP
Dossier for the IFS/BRC auditor	Manual reconstruction, days	Automatic per-batch dossier with curve and image

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 put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Plant with 1-2 flash pasteurization lines (plate or tubular heat exchanger), multi-SKU (orange juice, multifruit, chilled freshly squeezed), variable CIP.
• Edge + Connect pilot (vision on the sight glass and the packing line + integration with the pasteurizer SCADA and the batch recipe). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how often cycles are repeated because of poorly documented CIP, scrap from oxidation caught late, and the energy saved by the optimal pair per SKU. A reasonable reduction of ≥30%, to be refined.
• Hard lever: one food safety alert avoided (from an undocumented CCP deviation) multiplies the ROI. The recurring saving in energy and cycles is the floor.

And the process manager's reasonable doubt

“What if the agent recommends dropping the setpoint by one degree and I end up with an alert?” — the agent does not sign setpoint changes: it proposes, and the proposal goes through the second validation ring before it reaches the process manager. In anchored tasks (calculating the effective time-temperature pair from a SCADA curve), the best models brought error below 1.5% ^[1]. And the three safety rings are designed precisely so that critical operations are only executed by a person, where it matters.

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