Digitalization for the semiconductor industry.

The plant already has an ERP. Production reality and the system of record diverge every day.

Buying another overlay does not fix that. Execution has to be designed for how the floor actually runs — genealogy, e-records, an audit trail that is still there a year later.

We have designed MES and digital programmes from scratch inside semiconductor manufacturing, as employees, not as a vendor implementing someone else's blueprint. That is the work we do for you.

A characterization order looks like one dataset. Inside it: scalars, curve families, scope waveforms, and a handful of parts taken to failure.

Labs store them as files, then add disk when queries go from seconds to minutes. One schema cannot hold those four shapes without lying about at least three.

Typed stores per shape — relational for parameters, arrays for I–V families, blob plus decimation for waveforms, experiment records for destructive work — plus a catalogue that says whether two runs are comparable, and a tiering policy for what stays fast.

Read why a box plot over twelve destroyed parts is the wrong chart →

The time between a tester finishing and an engineer having something usable is export, format wrangling, deskew bookkeeping, and paste into a template.

Every vendor writes a different file. None of them writes the one your store wants. Setup metadata — the only thing that makes two runs comparable — usually never leaves the bench.

Connect one path: tester, SMU, curve tracer, scope, or prober. Normalise what comes off it, including the setup, into the structure from 01.

The question in a qualification review is not the figure. It is which parts, under which conditions, with which exclusions, using which version of the extraction.

Most labs can answer that with a person who remembers. Few can answer it a year later, or regenerate the figure without a silent change.

Lot, wafer, die, site, condition — and versioning around it — so every released figure carries provenance. Regenerating it either reproduces the number or says exactly what changed. Same discipline as an auditable execution system, applied to measurement.

Nobody reads seventy curves carefully every time. Site-to-site offset, a distribution going bimodal, two parameters that normally track drifting apart — visible in the data, invisible in practice.

"This looks unusual" is worthless to someone who has to defend the number. A model that calls the anomaly without a test, an effect size, and a sample count is a liability.

Classical checks first. Language models rank and write the sentence. Every finding carries the evidence plot. Every check has an explicit not enough data to say.