Computer Vision
Computer vision pilots are evaluated on curated frames and deployed into real lighting, real occlusion, real camera drift and real dust on the lens. The accuracy number from the pilot is almost never the accuracy number you get, and the gap is a data problem long before it is a model problem.
The pilot-to-plant gap
A model trained on clean, well-lit, well-framed images meets a production line where the light changes through the day, the part arrives at an angle nobody photographed, and someone repositions the camera during maintenance. Accuracy degrades quietly, because unlike a crashed service a mis-classifying vision model keeps returning confident answers. The second constraint is annotation: for defect detection the interesting class is rare by definition, so the expensive work is not training but obtaining enough examples of the failure you care about. The third is deployment reality — the bandwidth to stream every frame to the cloud frequently does not exist, and the latency budget on a moving line does not permit a round trip.
The three calls that decide the outcome.
Made explicitly, with the trade-off written down, before anything gets built.
Edge or cloud inference
Cloud gives you larger models and simple updates. Edge gives you latency measured in milliseconds, resilience when connectivity drops, and — often decisively — the ability to process video without transmitting it anywhere. Bandwidth is usually the forcing constraint: continuously streaming multiple camera feeds is expensive where it is even possible. The common answer is a hybrid, with detection at the edge and only events, crops or embeddings sent onward.
How you get enough of the rare class
Defect detection is an imbalance problem before it is a vision problem. The options are targeted collection, aggressive augmentation of the examples you have, synthetic generation, or reframing the task as anomaly detection so the model learns normal and flags deviation rather than learning each defect type. Which one applies depends on whether your defects are enumerable and stable, and getting this wrong wastes an annotation budget.
Model size against the deployment envelope
The largest model that fits is rarely the right answer at the edge, where thermal limits, power and inference latency bound the choice. Quantisation and distillation typically recover most of the accuracy at a fraction of the cost, and the decision should be made against the operating envelope of the actual device on the actual line — not against a benchmark table produced on a datacentre GPU.
The capabilities we bring on day one.
Each engagement assembles from this menu, sized to your scope, paced to your calendar.
Real-condition evaluation
Tested against the lighting, angles and occlusion of your actual site, not a curated validation set.
Annotation strategy
A plan for obtaining the rare class before the budget goes into labelling the common one.
Edge deployment
Models sized to the real thermal, power and latency envelope of the device on the line.
Privacy by construction
Anonymous analysis where identification is neither needed nor lawful, decided at design time.
Privacy as an architectural decision
Vision systems in public or workplace settings sit close to the sharpest edges of the EU AI Act. Real-time remote biometric identification in publicly accessible spaces is prohibited outside narrow exceptions, and emotion inference in workplaces and educational settings is likewise restricted — so the difference between counting people and identifying them is a legal boundary, not a design preference. The billboard platform we built is a deliberate example: it estimates audience size, general age group, crowd density and time of day to select an advertisement, and stores no facial identities or personally identifiable information. Anonymous-by-construction is both the compliant position and the one that survives scrutiny.
From kickoff to live in four phases.
Each phase has named deliverables, named owners and a named gate, and every one respects the systems you already have in place.
Discover
Site conditions, camera placement, latency and bandwidth reality, and an honest audit of what imagery exists versus what is needed.
Design
Task framing, annotation and data-collection strategy, edge-versus-cloud split, and the privacy boundary set before any model work.
Build
Data pipeline, model iteration against real-condition holdouts, and integration with the systems that act on the output.
Run & hand off
Monitoring for camera drift and distribution shift, retraining triggers, and enablement for the site team who will live with it.
What a good fit looks like.
Stated up front, so neither of us spends a call finding out this was the wrong conversation.
Engagement shape
A scoped sprint of roughly eight weeks, usually starting on a single line, site or camera cluster before any rollout decision.
From your side
Representative imagery from the real environment, site access or a detailed walkthrough, and a named owner at the site who will act on the output.
When not to hire us
If you need identification of specific individuals in a public space, we are not the right partner — that is prohibited or high-risk under the EU AI Act outside narrow exceptions, and we will not scope it.
Computer Vision questions, answered.
Less than most teams expect for common classes and more for rare ones, which is where defect detection actually lives. The plan depends on whether your defect types are enumerable and stable — if they are, targeted collection and augmentation work; if new failure modes keep appearing, anomaly detection that learns normal and flags deviation is usually the better framing.
Only if latency, connectivity or bandwidth force it — and one of them usually does, since continuously streaming multiple camera feeds is expensive where it is possible at all. The common design is hybrid: detection at the edge, with only events, crops or embeddings sent onward for storage and analysis.
In most cases yes, and it is usually the right design. Counting people, estimating crowd density or reading a general age band requires no identification and no retention of images. We built exactly that for the smart billboard platform, which selects advertisements from anonymous audience context and stores no facial identities or personally identifiable information.
Almost always because the evaluation set did not contain the conditions the floor produces — changing light through the day, unusual part angles, occlusion, camera repositioning after maintenance. A vision model does not crash when this happens; it keeps returning confident wrong answers. We evaluate against real-condition holdouts and monitor for drift precisely because the failure is silent.