The most common question we hear before starting a project is: "how do we know AI will actually detect our defect before we invest in a full production station?" The answer is a feasibility pilot — time-bounded, run on real data, ending in a clear recommendation.
Step 1 — initial consultation
We start by defining the product, the specific defect, and imaging conditions — camera, lighting, positioning. We also establish the business value of the problem: what a missed defect or a false rejection currently costs.
Step 2 — data
We prepare representative OK and NOK samples plus an independent test set the model never sees during training. Data quality and diversity at this stage determine how trustworthy the whole pilot result is.
Step 3 — model
The model comes from automatic training in AIVQC Trainer — with no limit on attempts — and we select decision thresholds matched to process risk — a false accept on a critical defect is weighed differently than a cosmetic deviation with no functional impact.
Step 4 — test
We measure detection quality, false-accept and false-reject rates, and performance — decision time on the target hardware. This is where numbers replace assumptions.
Step 5 — decision
You receive a qualification report with a clear recommendation: GO, CONDITIONAL GO, or NO-GO. An honest, data-backed NO-GO is just as valuable as a GO — it saves the cost of a misdirected investment in a full production station.
Scope and timeline
- Duration: 30 days, with a free extension available
- Scope: one Trainer workstation, one inspection station and up to three defect classes
- Format: with our support, or on your own with your team
- Pricing: quoted individually — a pilot and deployment with us, or the software alone if your own team deploys it
After a successful pilot, we prepare a separate production deployment offer covering licensing, hardware qualification, station configuration and an agreed scope of line integration.