OmniVision AI

Hardware

One box. The whole watchtower.

The EdgeBox is our blueprint for a purpose-built OmniVision appliance. It runs detection, captions, search and chat locally, plugs four cameras straight in, and is built from standard parts any local shop can replace.

Design goals

Built for a camera site

One box per site

Portal, AI models, workers and recordings all run locally. Only remote access uses the internet.

Cameras plug in

Four PoE+ ports power the cameras and keep them on their own network, with no route to the internet.

Quiet and cool

The recommended GPUs draw 70 W from the slot, with no extra power cable, in a 15.9-litre case.

Comes back on its own

Powers on when mains returns, and the software restarts itself. No one needs to visit.

Tiers

Same box. Choose your brain.

All three tiers share the same parts; only the graphics card changes. More GPU memory means bigger AI models and more detailed captions, not more cameras.

EdgeBox S

8 GB GPU · RTX 5060

  • Today's captioning model, fast
  • Search, alerts, counting and telemetry
  • Ask OmniVision runs, sharing the GPU
  • Best for smaller sites

EdgeBox X

24 GB GPU · RTX PRO 4000

  • Larger models with headroom
  • Ready for future, bigger captioning models
  • 70 W, compact SFF card
  • Best for large or demanding sites

Connections

Every port has a job

Cameras go on their own bridged network behind the box. The internet port is only for remote access, cloud imports and alerts, through an encrypted tunnel with no ports opened on your router.

More than four cameras? Add a PoE+ switch to grow to 12 or more.

PortWhat it's for
CAM 1–4
PoE+, 1 Gb/s
IP cameras: video and power over one cable
WAN
2.5 Gb/s
Remote access, cloud imports, push alerts, model downloads
USB 3
rear
UPS signal for clean shutdown, SD reader for dashcams
USB-C
front
Import footage from USB drives
HDMI / DPLocal monitor or guard-station video wall

In development

Keeps watching when the power doesn't

For outage-prone areas, we're designing battery backup in two layers. It is a concept today, and the figures below are design targets, not measurements.

Built-in battery

Up to 100 Wh inside the case. Rides through flickers and short cuts, and always shuts the box down cleanly. Planned for every tier.

Battery base · P-R

A larger LiFePO4 base docks underneath for overnight outages, with an outage mode that saves power. Can charge from solar.

Your own hardware

Prefer a mini PC?

OmniVision also runs on off-the-shelf machines, on Windows or Linux. We publish a hardware guide with tested requirements and recommended models for every budget.

  • No GPU: everything except AI captions and chat, or cloud-assisted captions
  • 8 GB NVIDIA GPU: local captions
  • 12–16 GB: local captions and Ask OmniVision together
  • Windows installer, or Docker on Linux
PartMinimumRecommended
RAM16 GB32 GB
GPUnone, or 6 GB12–16 GB NVIDIA
System disk256 GB NVMe500 GB NVMe
RecordingsAbout 1–3 GB per camera per day
OSWindows 11 or Linux (Ubuntu 24.04 LTS)

Not sure what your site needs?

Tell us how many cameras you have and what your network looks like. We'll recommend the right tier, or tell you if the hardware you already own will do.