Computer vision measured on the defects it misses
Centricone Technologies builds visual inspection and detection systems on your own footage: what counts as a defect, how capture and lighting have to be set up, what the model scores against, and what happens to the cases it is not sure about.
Every vision system trades false alarms against missed defects. That trade is a business decision, not a technical default, and we make it with you before the line depends on it.
support coverage across US and Canada time zones
of code reviewed, tested, and documented before release
The standard lives in people's heads
Inspection depends on who is on shift
The threshold drifts between operators, between sites, and between the start and the end of a long day — and nobody can point at where it is written down.
The proof of concept never met the real line
It worked on clean images in good light, and the first week of real conditions — glare, dust, a new supplier's packaging — ended it.
Nobody has said which error costs more
A missed defect and a false alarm have very different prices, and until someone says which, the system is tuned to a number that suits neither.
for teams inspecting more than people can
Centricone covers the whole system: capture and conditions, the labelled set, the model and its thresholds, the operator interface, and the monitoring that catches the day the line changes.
Deliverables
What you get
- A running inspection system on your line
- A written defect definition, agreed with operators
- A labelled image set, kept as your standard
- Thresholds set against the cost of each error
- An operator review and override interface
- Drift and flag-rate monitoring
How a vision engagement runs
Agree the defect and see the line
What counts, and the real conditions the system will work in. This step, not the modelling, decides whether the project can succeed at all.
Capture and label
Footage from the actual line, labelled with your operators, with the held-out set reserved before any model is fitted to it.
Model, threshold, and trial
Scored against the held-out set, thresholds set to your cost of error, then run alongside people rather than instead of them.
Hand over with the monitoring
Operator interface, drift alerting, and an agreed process for re-labelling when the line changes.
When computer vision is the wrong build
Vision projects fail on the inputs far more often than on the model, and the failure is usually visible before anyone writes code.
- Your own operators cannot agree on what a defect is. Settle that first — no model can be more consistent than its labels.
- The images do not exist and cannot be captured. Vision starts with capture, and there is no way around that.
- The task is reading text and fields off documents. That is document intelligence, and it is a different build.
- The volume is small enough that a person can look at all of it. Then a person should.
Why teams build vision with Centricone
The defect defined in writing before any model
Built on footage from your real line
Thresholds set to the cost of each kind of error
Operators kept in the loop, not designed out
The capability behind this engagement
Model selection and evaluation, data pipelines, guardrails, and the monitoring and cost controls behind a system like this — covered in more depth on the AI development services page.
Frequently asked questions
Not sure this is the right shape? Tell us the situation and we’ll say which engagement fits.