Why most computer vision pilots never reach production
Lighting, camera placement and label quality decide more outcomes than model architecture. A field checklist you can run before spending on GPUs.
What we have learned shipping AI and software into real operations — written down, free to read.
Home › Resources
Everything here comes out of a real project. If it did not survive production, it is not on this page.
Lighting, camera placement and label quality decide more outcomes than model architecture. A field checklist you can run before spending on GPUs.
Latency budgets, bandwidth cost and data-residency rules compared side by side, with a simple decision table.
How to size a pilot, what to measure, and when to stop — written for non-technical sponsors who have to defend the spend.
Integration patterns that work when there is no modern API, and the ones that quietly create technical debt.
Three lightweight checks that catch most degradation, and how to wire them into ordinary alerting.
What you can measure without storing identities, and how to document it for a compliance review.
Overviews you can forward internally, plus the technical datasheets for the people who will ask the hard questions. No form wall — click and the PDF downloads.
Yes — that is how most engagements start. A time-boxed pilot on your real data, at a fixed fee, with a clear go / no-go decision at the end.
Usually not. Most existing IP camera setups are good enough. If placement or lighting is the limiting factor we will say so during discovery rather than after.
Under an NDA, on infrastructure you approve, with least-privilege access and deletion at the end of the engagement unless you ask otherwise.
That is the intent. We hand over documented code, a runbook and a training session. Many clients take over completely; others keep us on a support plan.
Tell us the topic. If we have hit that problem in a project, we will write it up and send it across.