Cameras that count,
video that never leaves.
A retail chain in Saudi Arabia already had CCTV in every store, and no operational data coming out of it. I led the AI and edge work that turns those existing cameras into live operations data, processed on a box inside the store so no raw video ever leaves the premises.
01The constraint that shaped everything
Saudi PDPL treats video of identifiable people as personal data, and blurring faces does not count as true anonymization: a blur is a transformation, not a guarantee, and the original is still the thing you transmitted. So the raw frames could not go anywhere. Not to a cloud GPU, not to a regional endpoint, not into a bucket for later.
That single fact inverts the usual build. Instead of moving the video to the compute, the compute moves into the store, and the only thing allowed out is the conclusion.
"Blurring a face is not anonymization. Once that is true, the video cannot leave the building, and every other decision follows from there."
02What I led
I owned the AI and edge side end to end: choosing what runs on-device, getting it to run there within the thermal and throughput budget of a small box, and deciding which signals were reliable enough to put in front of an operations team.
Two calls I would point to. First, the model licensing: I moved the detector onto a permissively licensed model, because several of the obvious choices carry copyleft terms that follow you into a commercial product, and a licensing problem found after rollout is not one you can patch. Second, a measurement bug that was quietly gating three of the analytics features: the counting logic looked finished and was wrong in a way that only shows up against real footage. Finding it took reading behaviour rather than reading code, and the correction unblocked all three at once.
Deliberately light on architecture here. It is a live commercial product, so the design stays with the company.
03Where it stands
Pilot stage, running on customer hardware. No accuracy or throughput numbers on this page on purpose: they have not been measured yet, and a number invented for a portfolio is worth less than admitting it does not exist.
On what a Jetson can and cannot actually do once DeepStream and TensorRT are involved: What a $250 Box Can Actually Do.