Industrial vision
Count, track and flag what the cameras already see.
Cameras nobody has the hours to watch
The cameras are already installed and already recording, and nobody has the hours to watch them. The counting, the checking and the exception-spotting still happen by hand, or not at all.
Counting, tracking and flagging, measured on your footage
Monitoring against what the footage actually supports: counting, tracking, dwell and flow, and flagging the events you care about — on a fixed camera, at the frame rates real operational video runs at.
We are deliberately conservative about accuracy claims. Detection quality depends on your lighting, your angles and your scene far more than on any published benchmark, so the pilot measures it on your footage before anyone commits to a rollout.
What the vision engagement leaves behind
- Measured accuracy on your own video, stated with its failure cases, not a vendor benchmark.
- Throughput and hardware sizing, so the running cost is known before you scale it.
- Exception alerts that route to a person who can act, rather than a dashboard nobody opens.
The honest limit on this work, and the people we will not identify
Honest limits: fixed cameras are our strong case — a moving camera invalidates the background model by construction. And we will not identify individuals; this is counting and flow, not surveillance of named people.
If you already use the hosted API: tracking and overlap suppression run on standard (CPU) infrastructure today; raw-frame object detection is served from a GPU-backed endpoint, sized and priced separately. This engagement is for a bespoke build beyond what the hosted verbs cover, not a replacement for them.
Where this starts
A week-long diagnostic, fixed fee, ending in a written assessment that names the decision, the measurable outcome and whether this is worth building at all.