What changed between shot forty and shot fifty-five

The problem

At tryout, what is missing is not the eye. It is memory and correlation.

Every tryout loop costs press time, steel, bench hours and calendar. And what decides whether another loop is needed is one person's reading of a handful of parts, at the end of a long day. When the defect sits on a Class A surface — the one the car buyer will see — the tolerance is zero and the discussion with the customer is long, because neither side has structured evidence: only impressions and loose photos. What is missing at a tryout is not the eye. It is memory and correlation.

• Tryout is the moment of truth: the mold is mounted, shot, and the part inspected. A person with years of craft reads it shot by shot for hours, noting on a sheet.
• They get it right, but they cannot correlate a hundred shots with a hundred parameter combinations, and what decides whether another loop is needed is their reading of a handful of parts at the end of a long day.
• Every loop costs press time, steel, bench hours and calendar on the critical path.
• On a Class A surface the tolerance is zero and the discussion with the customer is long, because neither side has structured evidence: only impressions and loose photos.

How it fits the IRIS system

Edge — AI vision at the press itself, with local inference and nothing leaving the plant.

Edge: AI vision at the press itself, with local inference and nothing leaving the plant.

The camera classifies each part in milliseconds against the shop's real defect catalog — sink marks, flash, flow lines, gas burn, ejector marks, warpage, short shots — and stores the verdict next to that shot's pressure, times and temperatures and the suspect detail: venting, cooling, insert. The technician stays in command: the AI gives the pattern, he decides the adjustment. What changed between shot forty and shot fifty-five stops being a question asked from memory at the end of the day.

• An industrial camera over the ejection area of the tryout press.
• A vision model trained on good parts and on the shop's real defect catalog: sink marks, flash, flow lines, gas burn, ejector marks, warpage, short shots.
• Every classification is stored next to that shot's parameters (pressure, times, temperatures) and to the suspect mold detail: venting, cooling, insert.
• The tryout technician stays in command: the AI gives the pattern, he decides the adjustment.

The cumulative effect is a history that lets the shop start the next similar mold knowing what happened on the last one.

See the full IRIS architecture →

Before and after

Today's tryout versus the tryout with Edge

Aspect	Today	With iLEAN Edge
Record of each shot	A handwritten sheet of impressions	Defect per shot, tied to parameters
What changed between shot 40 and 55	Nobody can say	The parameter and the defect, side by side
Root cause	By experience	With evidence, pointing at the suspect detail
Discussion with the customer	Impressions and loose photos	Structured evidence per shot
Loops per complex mold	The historical number	At least one fewer
Next similar mold	Starts from zero	Starts knowing what happened on the last one

from a handwritten sheet of impressions → defect history per shot correlated with parameters. From diagnosing root cause by experience → reaching it with evidence, and taking that evidence into the customer meeting.

Impact estimate

Impact estimate — to be validated against your loops-per-mold history.

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.

• An estimated reduction of at least one tryout loop on complex molds.
• Estimated payback 5-11 months.
• Every loop avoided frees press, bench and calendar days on the critical path.
• And the customer meeting on a Class A defect is held with evidence per shot instead of impressions.

an estimated reduction of at least one tryout loop on complex molds; estimated payback 5-11 months. Every loop avoided frees press, bench and calendar days on the critical path. *Estimate to validate against the loops-per-mold history*.

And the fair question from the production manager

"Will it call a good part defective and send us chasing a ghost?" — the false positive is the real risk of any vision system, which is why the model is trained on good parts and on the shop's own defect catalog, under that press's lighting, not on a generic model. Borderline shots are not silently classified: they are escalated for the tryout technician to decide, and every decision feeds back into training. The model proposes the pattern; the adjustment is always his [1].

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