Book a call ↗

Illustrative video operations room with security camera feeds and an edge server

Project 01 / Air-gapped video AI

Video
surveillance.

For an a16z-backed security-technology company.

← All our work

The project

Video at the centre.
Infrastructure underneath.

Client
An a16z-backed security-technology company
Our work
Video surveillance
Environment
Air-gapped, inside the customer site
Technology
NVIDIA L40S · Kubernetes · MediaMTX
Across the stack
Infrastructure · Applied AI

The service model

From camera feeds
to a shared view.

Inside the customer site, live streams pass through MediaMTX to AI inference on shared GPUs.

Camera feedsLive streams
Video platformMediaMTX · Shared GPUs
Monitoring teamsOperational visibility
Simplified service model

Engineering spotlight / Air-gapped NVIDIA

Live video AI.
No internet required.

An air-gapped Kubernetes platform that runs video AI on shared NVIDIA GPUs inside the customer site, with every install and update delivered offline.

25–40%

Faster incident response

Video AI runs locally, inside the customer site.

40–60%

Less deployment time

Helm-standardised rollouts from an offline registry.

30–50%

Better GPU utilisation

GPU slicing and time-slicing share each card across streams.

Reported results for this engagement, also covered in the video intelligence case study.

The challenge

Three constraints.
One closed site.

Surveillance AI had to run where the cameras are, with no internet connection.

01 / Offline

No internet.

Video AI had to run fully offline, inside the customer site.

02 / Compute

Idle GPUs.

Expensive GPUs sat idle between streams.

03 / Releases

Slow releases.

Releases into the air-gapped site took days.

The operating stack

Every layer,
inside the perimeter.

Built to install, update and run with no internet connection. Observability and security run across the platform.

Video applications

Video AI services · Monitoring views

Stream routing

MediaMTX, scaled horizontally for live streams

Model serving

Inference workloads on shared GPUs · Helm charts

GPU enablement

NVIDIA GPU Operator · GPU slicing · Time-slicing · DCGM

Kubernetes platform

Air-gapped Kubernetes · Offline registry

Hardware

NVIDIA GPU servers inside the customer site

The fleet

Inside the perimeter.
Nothing outside it.

320 live camera streams on 32 NVIDIA GPUs, run by a platform that has never been connected to the internet.

Control plane / 3 nodes

Platform and registry.

Kubernetes control plane, the offline registry, Prometheus and Grafana.

Stream tier / 6 replicas

MediaMTX.

Ingests and routes 320 live streams, scaled horizontally.

GPU pool / 8 servers

32 NVIDIA L40S.

Sliced and time-shared, running detection and tracking models across every stream.

Observability

DCGM · Prometheus · Grafana

Live telemetry on stream health and GPU load.

Alert rules and runbooks written for a site with no internet connection.

Security

Air gap · Offline registry · Kubernetes RBAC

Images and charts reach the site only through the offline registry. Nothing is pulled from the internet.

Offline delivery

Every update crosses
the air gap the same way.

Installs and updates arrive as one package, load into the local registry and roll out with Helm.

  1. 01

    Build

    Images and Helm charts

  2. 02

    Scan

    Checked before export

  3. 03

    Package

    Offline bundle

  4. 04

    Transfer

    Into the perimeter

  5. 05

    Load

    Offline registry

  6. 06

    Roll out

    Standard Helm release

  7. 07

    Validate

    Live streams and GPU load

  8. 08

    Operate

    DCGM telemetry

How GPUs are shared between streams

The NVIDIA GPU Operator manages drivers and device plugins on every GPU node. GPU slicing and time-slicing let several video workloads share one card, so capacity no longer sits idle between streams.

How live streams scale

MediaMTX runs as a horizontally scaled service on Kubernetes. More streams mean more replicas, not a larger single server.

How the platform is observed offline

DCGM exporters send GPU metrics to Prometheus, and Grafana shows them beside stream health. Telemetry, dashboards and alerts all stay inside the perimeter.

The outcome

Local AI.
Better use of every GPU.

The same engagement is told from the client’s side in our video intelligence case study.

01 / Local AI

Video AI inside the site.

Streams, models and results stay within the customer’s perimeter.

02 / Shared GPUs

Capacity in use.

Slicing and time-slicing turned idle cards into shared capacity.

03 / Offline releases

Updates as routine.

A standard Helm rollout through the offline registry replaced releases that took days.

← All our work