Ariel Guidi

← All experience

Nhood · Retail real estate

Video that never leaves the store

Hybrid edge-to-cloud architecture, with NVIDIA's engineering teams

2025

Video leaving site
None
Outcome
Production
Afterwards
Reusable offering

Nhood needed footfall analytics without sending identifiable video to the cloud. I designed the hybrid split and network path, then worked with NVIDIA's engineers to adapt the multi-camera tracking stack for Jetson hardware on site.

Why this problem exists

Nhood leases retail space, and the price of a location depends on how many people pass through it and where they linger. Measuring that well turns a negotiation based on assertion into one based on data.

The measurement requires cameras, and cameras in a shopping centre record identifiable people. Under GDPR that makes the raw video the most sensitive asset in the system. So the constraint that shaped the entire architecture created the architectural boundary: analytics could reach the cloud, but the footage could not leave the site.

Moving cloud-designed tracking to the edge

NVIDIA Metropolis, including the multi-camera tracking workflow we needed, was designed to run in the cloud. Its modules assumed cloud-side execution. Running it beside the cameras required adapting those components to build and run on on-premises edge hardware.

We did that in collaboration with NVIDIA's own engineers, and adapted their MTMC multi-camera tracking workflow for hybrid operation: tracking across several camera views happens on site, and only the derived measurements are synchronised outward.

Cameras feed Jetson edge hardware inside the store, where multi-camera tracking runs. Only aggregated analytics cross the boundary into AWS for dashboards. Raw video never leaves the site. on site, client infrastructureCamerasraw videoJetson + MTMCtracking, on-premAWSAnalyticsaggregatesDashboardsmetrics only, over VLANvideo never crosses this boundary
The privacy property is structural rather than procedural: raw footage has nowhere to go, because the only path outward carries aggregates.

The network enforced the privacy boundary

The edge nodes needed a route for derived analytics without creating a path for raw footage. My responsibility covered the network design between the client's VLANs and AWS, restricted to the services required by the edge workflow. The resulting boundary made the privacy property structural: only aggregates had an outward path.

Production result

The architecture reached production and became a reusable internal offering. The implementation knowledge for running the cloud-designed vision stack on site was documented and packaged for later engagements.

Ownership boundary. NVIDIA's engineering teams worked on recompilation and optimisation. My role covered the architecture, hybrid split, networking, and production delivery.

Stack

  • NVIDIA Metropolis
  • NVIDIA MTMC
  • DeepStream
  • Triton
  • Jetson
  • AWS
  • Kubernetes
  • VLAN networking