Edge AI is the compute that sits where the guard was supposed to stand, which in practice is a few inches of HDMI cable between the DVR and the wall monitor in the leasing office.

Edge AI is machine learning inference that runs on hardware sitting at the same physical location as the data source, not in a cloud region. The textbook definition (NVIDIA, IBM, Red Hat, Cisco) stops there and then pivots to autonomous cars and smart speakers. In the security context the more useful framing is that the edge is wherever an attentive human would have watched. For a multifamily or commercial property that place is a wall monitor in the leasing office or lobby, driven by a DVR or NVR over HDMI. Edge AI in security is the compute that replaces the stare of the guard who was never actually going to watch that monitor for eight hours straight.

The canonical definition assumes the sensor and the inference live on the same device. A smart speaker is the microphone and the wake-word model. A smart camera is the lens and the person detector. That framing works for greenfield IoT deployments where you buy new hardware. It does not describe the existing installed base of physical security, which is tens of millions of analog HD-CVI or HD-TVI cameras wired to DVRs that know nothing about AI. In that world the edge is not the camera, it is a few inches of HDMI cable between the DVR and the monitor. Cyrano sits exactly there, which is why one $450 unit running at about 15 watts can make an entire existing camera system intelligent without touching the cameras or the network.

One HDMI input at 1920 by 1080 resolution, 30 frames per second. That is roughly 62 million pixels per second that the edge unit decodes, once, into a single frame buffer. Inside that frame the DVR has laid out up to 25 camera tiles in a multiview grid. The Cyrano unit auto-detects the grid (2x2, 3x3, 4x4, 5x5, or a custom layout) within about 800 milliseconds of HDMI signal lock, maps tile polygons, OCRs the per-tile camera labels from the DVR's on-screen display, and begins per-tile inference. Power draw is roughly 15 watts. Install is under two minutes. Cost is $450 hardware, $200 per month software. The edge is that unit.

Because on a real commercial property the RTSP streams are usually not reachable. Half the cameras are analog BNC on a hybrid DVR with no IP address. The IP cameras that exist are on a VLAN the property manager cannot change without calling the installer who set them up six years ago and is no longer in business. The NVR's ONVIF profile is from 2018 and authenticates in a way a modern edge device does not speak. The one place every camera feed reliably appears is the HDMI output driving the wall monitor. That signal is guaranteed to exist because a human was originally supposed to watch it. Defining the edge as that HDMI point sidesteps every RTSP and ONVIF integration problem at once.

Both are valid forms of edge AI. A smart IP camera with on-board analytics is edge AI in the strict sense, its inference runs on the camera itself. The problem with that form is that it requires buying new cameras. A Class B or Class C multifamily property already has 8 to 24 cameras installed, most bought between 2015 and 2022. Replacing them costs $50,000 to $150,000 per property. The HDMI edge answer brings edge AI to that existing system for $450 hardware per property and a 2-minute install. The two forms are not mutually exclusive. They target different buyers: smart-camera edge AI targets greenfield installs, HDMI edge AI targets brownfield retrofits.

It is accurate. The unit meets all of the technical criteria the canonical definitions list. Inference runs locally on silicon inside the device, on the same LAN as the source. Video never leaves the property. Alerts are produced with seconds of latency, not the hundreds of milliseconds an upload round-trip would add. When the internet uplink fails the device keeps running inference, queues alerts on disk, and flushes when connectivity returns. Nothing is routed through a cloud GPU. The only outbound payload is a short thumbnail plus event metadata. By the NVIDIA definition, by the IBM definition, by the Arm definition, it is edge AI. It is also a different physical placement of that edge than those definitions illustrate.

HDMI signal lock at about 800 milliseconds, EDID read from the source DVR, layout detection within the first second, tile polygon map produced, per-tile OCR of the DVR's on-screen camera labels in the next two to three seconds, model warm-start in parallel, first per-tile inference run complete by about five seconds, first real events surfaced to the dashboard within 10 to 20 seconds depending on motion in the scene. Within 90 seconds the unit has collected enough scene statistics to classify normal versus unusual motion for the loitering and restricted-area detectors. Within an hour the false-positive rate is at steady state. All on-device. No scene data uploaded to any training pipeline.

The uplink from the property carries only alert payloads. A typical alert is a sub-100KB JPEG thumbnail plus a small metadata blob (timestamp, tile label, event class, dwell seconds, track id). A busy multifamily property might generate 40 to 80 alerts per day, peaking during move-in and move-out. The total outbound data per property per day is typically under 10MB. Compare that to cloud-recorded video surveillance, where 16 cameras at 2Mbps each is 32Mbps sustained outbound, about 345GB per day. The bandwidth story alone is why properties with flaky uplinks can deploy edge AI and cannot deploy cloud recording.

The unit detects the layout change from the next frame, remaps the tile polygons, and keeps running per-tile inference against the new geometry without a reinitialization. During a fullscreen drill-in the entire frame becomes one tile and inference runs on that one tile at higher effective resolution. When the operator returns to multiview, the grid snaps back and polygons are re-mapped. No frames are lost to the swap. This is a real field behavior the generic edge AI articles cannot describe because they do not model the human operator who is still occasionally touching the DVR UI.

At one Class C multifamily property in Fort Worth, a single Cyrano edge unit caught 20 security incidents including a break-in attempt within the first 30 days of install. The property had 16 existing analog HD-CVI cameras wired to a hybrid DVR from 2019, a wall monitor in the leasing office that nobody watched in the evenings, and no on-site guard. The HDMI edge unit was the only change. Everything else (cameras, cabling, DVR, network) was already in place. That is the concrete shape of edge AI when the definition is 'compute where the guard was supposed to stand.'