The existing inspection camera, stitched to the ERP

The problem

The camera inspects every closure, and its data stays on its console.

The plant already has an in-line vision system inspecting every closure at high speed: thread, color, defect. But its data lives isolated in its own console. Rejection rate and defect codes stay on the vision equipment, never linked to the resin batch or the order running at that moment.

• The plant already has an in-line vision system checking every closure at high speed: thread, tamper-evident band, color, short shot, liner presence.
• Its data lives isolated in its own console. Reject rate and defect codes stay on the vision equipment and are never linked to the resin batch or the order running at that moment. The console shows today's reject rate; it does not know which bottler that batch is going to.
• When a bottler complains, someone exports logs by date, guesses which batch was running and matches them by hand. It takes days.
• The integration was always planned, but it never reached the top of the IT list because the camera works and the ERP works. The gap sits between them. Meanwhile, every claim costs the same days of manual matching.

How it fits the IRIS system

Connect stitches the vision console to the ERP through the API it already has.

Through the API the vision system already has, Connect automatically matches every piece-by-piece verdict with the active batch and order in the ERP. The rejection summary per batch becomes available in the ERP with nobody transcribing anything.

Nothing is replaced and no camera is reconfigured. Connect takes the verdicts through the interface the vision system already exposes and matches each one to the batch and order active in the ERP. The quality summary per batch is simply there when someone asks. Quality stops exporting logs and starts reading trends per mold and per resin batch. That is where the vision investment starts to pay twice.

See the full IRIS architecture →

Before and after

Inspection data today versus inspection data on the batch

Aspect	Today	With iLEAN Connect
Where closure reject data lives	The vision console	The ERP batch record
Link to resin batch and order	None	Automatic, per verdict
Defect codes per batch	Exported by hand	Summarized in real time
Answering a bottler's claim	Days	Minutes
Inspection recipes and thresholds	—	Unchanged
Integration project in the IT queue	Waiting	Not needed

Reject data isolated in a local console → real-time per-batch quality summary in the ERP.

Impact estimate

Impact estimate — to be validated against your recent bottler claims.

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 3-6 months, one of the fastest in the matrix. The camera and the ERP already exist; only the stitching is new.
• Customer claim response goes from days to minutes: the reject history of the batch in question is already in the ERP. The answer arrives while the bottler is still on the phone.
• Reject rates per batch can be compared across molds, colors and resin batches without exporting a single log. A mold whose reject rate creeps up shows before it becomes a claim.
• And the vision investment the plant already made starts producing traceability, not just ejected parts. Its data finally reaches the people who negotiate with the bottler.

Payback 3-6 months · claim response in minutes. Fast payback of 3 to 6 months, customer claim response in minutes instead of days. Estimate to validate.

And the fair question from the production manager

“Isn't a proper integration project cleaner?” — it is, the day it reaches the top of the list, and that is the problem: it is the classic case of a modern system left isolated because there was never time. Matching a verdict to a batch by line and timestamp is an anchored task, where the best models drop below 1.5% error [1]. This stitching replaces neither system and needs no shutdown window, so it does not compete in that queue. If a full integration comes later, nothing done here is lost.

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