01
Hit one number. Before it moves.
JUMO makes platinum temperature sensors. Each must hit an exact resistance.
02
The laser sets it too early.
Printing and selection shift the resistance afterwards. The correction has to predict that.
03
47,115 substrates. 2,737 graded orders.
Eight files that shrink stage by stage. One blind merge would have lied.
04
Every gap meant something.
294 substrates broke before sputtering. 8,000+ had no order. One ID read 9.22E+18.
05
Which correction?
Some sensors are trimmed twice. The target is the last correction in the clean room: NCTr.
06
R² 0.987. We did not ship it.
The first run looked brilliant, on the wrong target. We went back.
07
5× lower error.
RMSE 0.1044 → 0.0202. 99.63% inside ±0.1 tolerance.
08
73 sensor types → 12 groups.
A parser reads type, geometry and resistance straight from the part number.
09
Every answer is traceable.
Model version, AI label and inputs are saved with each recommendation. The operator decides.
10
My part.
Framing, the source inventory, join audits, trust controls, the part number parser, the NCTr fix and the end to end test.
11
The law caught up.
EU AI Act transparency rules apply from 2 August 2026. Our app already labelled every AI output.
12
Next: SHAP and drift.
An explanation for every recommendation, and an alert when the line changes.