Vision · Manufacturing
Visual quality inspection for manufacturing
In visual inspection a single accuracy number is misleading: lowering the miss rate raises false alarms, and false alarms are what actually stop the line. Zero-Shot tunes that balance to your line speed and the cost of each error type, building the system from data captured under factory lighting.
What we build
- 01Detection of surface defects, scratches, missing parts and assembly errors
- 02Label, barcode and expiry date verification
- 03Alignment and dimension checks fast enough for line speed
- 04Station setup including camera, lighting and enclosure
- 05A self-check that notices dirty lenses and lighting drift
How it runs
01
Collect defective parts
This is the hard part, because defects are rare. Defective parts are gathered across shifts, and deliberately produced samples are used where necessary.
02
Set the threshold against cost
The cost of a missed defect and the cost of a false alarm are written as numbers, and the threshold follows those numbers rather than instinct.
03
Run in parallel
For a period the system runs without stopping the line, only recording. Its decisions are compared against human inspection.
04
Hand over authority
Once the comparison holds, the system gets permission to stop the line — starting at one station.
What we measure
- Miss rate — defective parts that pass
- False alarm rate and the line downtime it causes
- Decision latency against line speed
- Accuracy variation across shifts and lighting conditions
- Change in customer complaints and returns
When this is the wrong fit
If a defect is hard to see even by eye, or parts arrive in a different position every time, mechanical fixturing comes first. And if defects are extremely rare — a few cases a year — strengthening existing inspection beats building a system.
Common questions
How many defective samples do you need?
A few hundred per defect type gives good results. You can start with a few dozen, but accuracy stays low and improves as samples accumulate. We plan sample collection in the first week, because it is usually the longest step of the project.
What if it stops the line by mistake?
That is why the system runs without authority first: it decides but does not intervene. Line-stop permission is granted only after the false alarm rate is measured and brought to an acceptable level, and it starts at a single station.
Do we have to rebuild it for a new product?
No, but samples for the new product are needed and the threshold is re-measured. That process is far shorter than the original build, and the system is designed from the start to accept new products.
Other work in this sector
Other services
What we've written on this
Last updated: 2026-09-05
Let's talk about this one
Tell us what you want to build and within two days we'll tell you who takes it and how we'd measure it.