Computer vision extracts information from images and video: reading text from documents and forms, inspecting products for defects, counting items, recognising faces or number plates, and monitoring safety conditions. It replaces visual checking that is repetitive, high volume and error-prone when done by tired humans.
Design for the wrong answer
No vision system is perfect, so the design question is what happens when it is wrong. A defect detector that silently passes bad units is dangerous; the same model with a review queue for low-confidence cases is genuinely useful. Accuracy targets should be set against the cost of each error type, not against a benchmark number.
The second reality is that model quality depends on training data resembling deployment conditions. Warehouse lighting, phone camera quality, crumpled documents — we collect from the real environment rather than an idealised one.
Process
How we deliver it
Process flow for Computer vision
01
Define error costs
What a false positive costs versus a false negative.
02
Collect real data
Images from the actual environment, including the bad ones.
03
Build and evaluate
Train or adapt a model, measure on held-out real cases.
04
Add the review queue
Low-confidence results routed to a human, not through.
05
Deploy and monitor
Watch for drift as conditions change.
Deliverables
What you receive
A deployed vision model meeting the agreed thresholds
Review interface for low-confidence cases
Accuracy report broken down by error type
Integration into your existing workflow or system
Monitoring, drift alerting and a retraining plan
Engagement shape
Eight to fourteen weeks including data collection, which is usually the longest phase and the one most often underestimated.
Stack decisions follow the problem. This is where we usually start, not a fixed menu.
Frequently asked
Questions we get about this
How many images do we need?
Fewer than the folklore suggests — a few hundred well-labelled, representative examples often outperforms thousands of near-duplicates. What matters is coverage of the awkward cases: poor lighting, unusual angles, damaged items. We would rather have 300 varied images than 5,000 identical ones.
Can it run without internet, on-site?
Yes. Edge deployment on local hardware suits factory floors, remote sites and anywhere latency or connectivity is a constraint. It costs more in hardware and adds an update process, and we size that honestly before you commit.
Talk it through before you commit
A discovery call is a working session on your constraint, not a sales pitch.
