A 1,400-camera video platform designed the obvious way — stream everything to a central GPU farm, archive all of it — hits a bandwidth and cost wall long before it hits a technical one. The architecture has to be economically survivable for three years, not just demonstrable once.
Architected the technical approach for a state-scale law-enforcement video platform: automatic number-plate recognition and face recognition across 1,400 cameras, with edge inference per camera rather than centralised GPU processing, event-triggered retention rather than full archival, and integration into national vehicle and identity databases.
Delivered the full cost and architecture model — three-year TCO, per-camera economics, and the build-versus-buy analysis — alongside a live six-camera reference deployment.