Products · #15 · Operations & Forecasting

TIER B BLUEPRINT

Industrial Computer Vision QC

Catch product defects at line speed - even defects you don't have many examples of. 100% inspection instead of sampling.

from €3,990

LINE CAM 2 - 100% INSPECTION · LIVE

SCAN

⚑ #4812 · NOVEL anomaly - not one of 7 known defect classes → engineer queue

Inspection log

#4810 ✓ pass ·· 41 ms

#4811 ✓ pass ·· 39 ms

#4812 ⛔ ANOMALY · 0.94

#4813 ✓ pass ·· 40 ms

#4814 ✓ pass ·· 43 ms

Defect evidence - #4812

edge chip · NE quadrant · 2.1 mm · heatmap + frame stored ⛓

NOVEL → anomaly path auto-reject ✓ line unaffected sample archived ✓

🟡 ENGINEER CONFIRMS - becomes defect class #8 and training data by Friday.

shift ▸ 4,812 inspected · 23 rejects · escape rate audited vs human QC

How it works

Optics first

Camera, lens and lighting engineered for your product and line speed. Half the accuracy is photons, not models - we say that upfront.

Learn normal, not just defects

Anomaly detection learns your good product. Rare defects get flagged from 3 examples - not 3,000 you don't have.

Decide at line speed

Pass or reject in under 50 ms on edge hardware next to the line. Every reject stores the frame, heatmap and score - where and why, not just "fail".

Close the loop weekly

Novel anomalies go to an engineer queue; confirmed ones become new defect classes. Escape rate audited against your human QC.

The Four Guarantees™ - this build

Measured value

Escape rate and false-reject rate agreed upfront against your current QC baseline; eval gate on a held-out defect set before the line goes live.

Defensible

Every reject carries frame, heatmap and score - batch-traceable evidence for customer claims and recalls, not a black-box verdict.

Self-correcting

Novel anomalies become new defect classes weekly; drift alerts on lighting and camera shift; false rejects reviewed to protect yield.

Yours & everywhere

Runs on edge hardware next to your line - no cloud round-trip. Full source. MCP endpoint for MES and QMS integration.

The number, sized honestly

Reference buyer: Lithuanian manufacturer - 1-5 lines, visual defects driving customer claims, manual QC sampling 5-10% of units today.

100% inspection - every unit, vs 5-10% manual sampling

<50 ms pass/reject decision at line speed, on the edge

2 rates escapes & false rejects - both agreed upfront (conservative)

~4 mo payback at the PoC tier at reference claim costs

Three ways to own it

Tier	What you get	Price
Scaffolding ★	The full repo - anomaly + classifier pipeline, edge inference stack, evidence store with heatmaps, eval harness, MCP server. Reference run on public industrial defect data.	€3,990
PoC ★★RECOMMENDED	Optics assessed and models trained on your product imagery - escape and false-reject rates measured on a held-out defect set from your line. Code + quality report yours. No guarantee before your photons prove it.	from €12,000
Implementation ★★★	Production: edge deployment on the line, MES/QMS integration, weekly novel-anomaly loop, lighting-drift monitoring, quarterly escape-rate audits - the agreed numbers (conservative) attach here.	from €36,000

★ = engagement depth. PoC is the recommended path: quality proven on your data before production money. The PoC carries no performance guarantee by design; the agreed number (conservative) attaches at Implementation, informed by the PoC report.

What we don't promise

Vision QC is only as good as its escape rate - so we measure it. Both failure directions are reported: escapes (cost you customers) and false rejects (cost you yield), audited against your human QC. Novel defects are flagged as anomalies rather than silently passed. But a defect invisible from the camera's angle stays invisible - lighting and optics are half the project, and that half comes first.

Ready to see your own number?

Request the build: within 48h you get a personal reply with the value sized to your volume.

No commitment · reply within 48h · your data stays in the EU