The questions you'd
ask in the meeting.

Yes, for safety and compliance work — PPE, restricted zones, proximity, smoke and spills, and that can be deployed in days without new hardware. No, not generally for defect inspection. A CCTV camera is the wrong resolution, the wrong lens, uncontrolled ambient light and no synchronisation with the part passing by, so the defect is often not in the image at all. No model can retrieve information that the sensor has never seen. We’d rather tell you that up front than sell you a pilot that fails quietly.

Good question and one we’d rather answer than duck. We are early — so the engagement is structured so you never have to take a big step on faith. A feasibility is a small, fixed scope piece of work that results in a written report. This includes the defects that we were not able to detect. The pilot is one line with acceptance criteria you agree in advance and runs in parallel to your existing QA, not replacing it. You own the images, the labels, and the model weights throughout, and the cell hardware is standard industrial equipment that another integrator could support. When you’re wrong about your application, you learn within weeks for a small amount of money — not year two.

Yes, we do. We design the cell, source the cameras, optics, illumination, enclosure and edge hardware, build and commission it and hand over a working system all under one commercial scope. During feasibility you get an itemised bill of materials so you won’t be surprised at anything about the cost. If you’d prefer to buy the hardware yourself to our specification, we’re happy to support that too.

So that’s why the feasibility phase is so important. You send out physical samples over good parts and every failure mode. We run imaging trials until the defect is reliably visible, train a model, and report precision and recall measured per defect class on a held-out set — not one headline percentage. If a defect class is not reliably detectable, that goes in the report as well. The acceptance criteria for the pilot are then written from those measured numbers and not from a sales promise.

How much depends almost entirely on the optics your defect needs — a presence/absence check and a sub-millimetre surface inspection are different orders of magnitude and you can’t honestly quote either from a website. But the commercial shape is consistent, with a one-time charge for the vision cell and commissioning, followed by an annual support-and-model-maintenance charge. At the end of feasibility, before any capital is committed, you will have an itemised bill of materials and full cost breakdown. Give us a call and on the first call we will give you a realistic range for your particular application.

For a stable, high-volume part, with well-defined geometric rules, a traditional rule-based system is often the right answer and we will say so. In cases of variable, cosmetic or hard to describe defects as a rule — stains, textures, weld appearance, organic shapes, mixed product — and where new variants would otherwise mean reprogramming every time, AI earns its place. We build both and the recommendation comes out of what is feasible rather than out of what we would prefer to sell.

New variants are done by collecting images and retraining, usually without touching the hardware if the part envelope is similar. Models are versioned, validated prior to release and rollback capabilities are built in. Under a managed service we look out for drift and we take care of retraining as part of the SLA. Under a capability transfer we train your engineers to do it, and you keep the tooling and the weights.

Yes. Both inference and reject decision are running on the edge device at the cell, so there is no dependency on the cloud in the control path. If the platform or WAN is unreachable the cell continues to inspect and reject, buffering its records locally and syncing when the link is restored. Full air-gap deployments supported.

Safety detection doesn’t have to identify anybody. Models operate on anonymised inference, on-device face blurring can be turned on before any frame is saved, retention periods are adjustable, and access to evidence clips is role-restricted and logged. We are happy to participate in that review directly with your HR and legal teams.

Annual support contract covering monitoring of model performance, retraining, platform updates, hardware spares and defined response times. Remote support and on-site attendance as required by the contract. Or we do a build-operate-transfer engagement, where we operate the system while your team learns and then we transfer ownership of the system, the models and the documentation.