Technology

Technology

More AI compute than a typical camera. Less exposure than cloud inference. Camera images stay on a private local network. This page is the technical design we are building — not a shipped spec sheet.

In development. First home pilot being prepared. Camera counts, UPS runtime, and model names will be published only after they are measured.

Local AI computer illustration

The three common architectures

ArchitectureAdvantageTrade-off
In-camera AIPrivate, low latencyVery limited compute budget
Cloud AILarge models, easy scalingImagery or derived data leaves the premises; vendor and internet dependency
Ilonai local AI applianceMuch more compute than a typical smart camera, imagery stays localRequires a local compute box and a dedicated network

That middle ground is the technical story. For a general audience we call it context-aware local AI. For technical readers, if this remains accurate as we ship: a fine-tuned multimodal / vision-language model on dedicated local hardware. We do not put model brand names on the customer homepage.

System diagram

System diagram: approved PoE cameras connect to a managed PoE switch, then to a local AI computer. The computer handles reasoning, a short buffer, fall decision, and health. Independent 5G carries alerts only.
Approved PoE cameras on a private wired network (RTSP / ONVIF). The local AI computer handles context-aware visual reasoning, a short encrypted rolling buffer, fall decision, camera and node health, and local encrypted incident storage. Independent 5G/eSIM carries heartbeat and emergency alerts only. Optional UPS sits on the compute side.

Your home or facility Wi-Fi does not need to become a surveillance network. Cameras are not meant to sit on ordinary household Wi-Fi or to require a camera-vendor cloud account.

Why centralized local compute?

More AI capability than a typical camera, less privacy exposure than cloud inference.

Putting a tiny detector in every camera is convenient and private, but the model has to fit a small processor. Sending video to a large cloud model is capable and operationally easy, and it moves sensitive imagery off site.

One (or a few) local AI nodes can serve many inexpensive cameras. Intelligence lives on the box; the camera can remain a replaceable sensor. We will publish home-node and facility-node camera ranges only after decoder, network, and model benchmarks exist. We will not promise “50 cameras easily” in advance.

For facilities, multiple compute nodes are the intended failure domain: one hardware fault should not disable an entire building.

Cameras: standards-based, approved list

We are not building a proprietary camera cloud.

Intended properties:

  • approved PoE network cameras
  • local IP networking
  • standard local video protocols (RTSP / ONVIF class)
  • cameras replaceable without rebuilding the AI system
  • no camera-vendor cloud required for core detection

Benefits we are designing for: lower per-room hardware cost at scale, simpler replacement, less lock-in, facility-standard cabling, and the ability to keep the camera “dumb” while Ilonai owns the decision.

We will qualify models for power-cycle recovery, RTSP reconnects, night mode, image quality, and stability. We do not promise arbitrary third-party camera compatibility.

Do not read “smart camera” here. That phrase usually implies a vendor app, cloud account, and remote viewing. Ilonai’s direction is the opposite.

Isolated network: cameras to a private PoE switch, then the local AI computer, then independent 5G for heartbeat and alerts only.
Cameras stay on PoE and a private switch. The local AI computer never needs household Wi-Fi. Independent 5G/eSIM is for heartbeat and emergency alerts, not a camera cloud.

Why this matters:

  • less dependence on customer IT
  • predictable networking
  • reduced attack surface versus cameras on everyday Wi-Fi
  • no camera access from normal household or office devices
  • independent cellular alerts
  • simpler install: no port forwarding, no home Wi-Fi credentials for the cameras

If true at launch: core detection still runs locally even if the external connection is unavailable.

Rolling buffer and incident storage

Video is not retained as a diary of the day. A short encrypted rolling buffer (about 10–15 seconds) exists locally and is continuously overwritten. If a possible fall is detected, that window can be frozen, plus a short post-event period, as a local incident clip.

Access is meant to be authorized and logged. Retention is meant to be short. Ordinary footage is not uploaded to a video cloud.

Heartbeat and failure detection

A safety system has two jobs: detect the emergency, and detect when itself cannot detect the emergency.

Designed monitors:

  • camera still delivering valid frames
  • AI computer health / heartbeat
  • missed heartbeats (lost cellular path)
  • mains power lost for more than a defined period (example: five minutes) → notify
  • recovery reported when power returns

We do not promise a UPS runtime until the hardware package is fixed.

Ilonai can know: a camera is offline, a heartbeat was missed, software version.
Ilonai cannot know remotely: what the person is doing in the room.

See also How it works, Privacy & Trust, and Evidence.