Four AI/ML products, one operating model — clinical workflows, visual inspection, data center operations and nonprofit CRM. Explore the portfolio

INSAIT Solutions
Request a demo

Evidence · InspectionAI

Real output, and the claims we refuse to make.

Unretouched frames from a model in production — including the low-confidence ones — the labelling standard behind them, and what it will not do.

Send sample parts

01 — Unretouched model output

What it draws, at the confidence it drew it.

23% and 25% are shown on purpose: a tool that only displays confident calls is not telling you how it behaves.

Scratch Neck 82% — Grey bench, even light
Scratch Neck 82%Grey bench, even light
Thread Damage 23% · 25% — Near-dark frame, corrected before it was judged
Thread Damage 23% · 25%Near-dark frame, corrected before it was judged
Scratch Head 21% · Scratch Neck 64% — Two classes on one part
Scratch Head 21% · Scratch Neck 64%Two classes on one part
Scratch Neck 70% · 77% — Rough concrete ground
Scratch Neck 70% · 77%Rough concrete ground
Thread Damage 35% · 28% — Two findings along one thread run
Thread Damage 35% · 28%Two findings along one thread run
Scratch Head 60% — Bright, reflective ground
Scratch Head 60%Bright, reflective ground

02 — Same model, same settings

Three grounds it was never retuned for.

PASS — White ground
PASSWhite ground
PASS — Near-black ground
PASSNear-black ground
PASS — Concrete ground
PASSConcrete ground

03 — How it was taught

A model is only as good as the boxes it was shown.

The rejected frames matter more than the accepted ones.

Accepted placement

Head damage — Box on the damaged area, not the whole head.
Head damageBox on the damaged area, not the whole head.
Neck scratch — Tight to the mark above the thread run.
Neck scratchTight to the mark above the thread run.
Thread damage — On the deformed threads themselves.
Thread damageOn the deformed threads themselves.

Rejected placement

Rejected — mostly background — More bench than part, so the model learns the bench.
Rejected — mostly backgroundMore bench than part, so the model learns the bench.
Rejected — whole head boxed — Teaches the shape of a head, not a scratch.
Rejected — whole head boxedTeaches the shape of a head, not a scratch.

04 — Boundaries

What it will not do.

Stated here rather than discovered in month two.

  1. A model only names the classes it was trained onThe class list is finite and defined up front.
  2. Two domains are deployed; the rest are trained in a pilotFastener inspection and adhesive peel analysis run today.
  3. An underperforming class is suppressed, on purposeOne fastener class fired on good parts, so it is switched off until retrained.
  4. It flags parts. It does not reject themNo actuator, no PLC write, no bin gate.
  5. It reports capability, not dimensionsDPMO, Z-bench and Ppk — not arithmetic on a measurement nobody took.
  6. Inconclusive is not a passAn empty frame is excluded from pass-rate arithmetic.

The number we do not publish

There is no accuracy figure anywhere on this site.

We have not measured one on your parts, and a number measured on ours would tell you nothing about yours.

Talk to INSAIT

The only number that matters is the one from your line.

Send sample parts