AI at the edge has two form factors. Every explainer on the SERP only describes one of them.

On the current SERP, it means machine-learning inference that runs on a device rather than in a cloud datacenter, almost always on new silicon shipped inside a new sensor (a smart camera, a lidar, a robot, a phone, a car). The NVIDIA, IBM, Cisco, Dell, Red Hat, SUSE, STMicroelectronics, and Scale Computing explainers all frame the category that way. The framing is correct as far as it goes, but it describes exactly one deployment pattern: greenfield, with the intelligence baked into a device you buy new.

A side-channel edge AI box that retrofits intelligence onto hardware already on the wall. Cyrano is one example. It is a physical unit that plugs into the HDMI output of an existing DVR or NVR, reads the composite multiview frame the DVR is already driving out to the guard monitor, and runs inference locally on the Cyrano device itself. No new cameras, no RTSP credentials, no ONVIF handshake, no cloud round trip. The cameras, the cabling, and the recorder do not change; only the HDMI line now has a witness on it that happens to be a neural network.

Because the DVR has already composited the streams for you. A 16-camera property's DVR outputs a single 1920x1080 HDMI frame with 16 tiles arranged in a 4x4 grid (or 5x5, or 1x1 when a guard drills into a single camera). Cyrano treats that composite as the input. One frame, one forward pass, up to 25 tiles resolved at once, then a per-tile spatial map back to camera names. You avoid opening 25 RTSP sessions and you avoid running the model 25 times. You also avoid 25 chances for a camera credential to be wrong.

A DVR burns graphics into its HDMI output that are not part of any camera feed: a live clock in the top-right corner, a per-tile name strip, a recording indicator. If you run person detection on that composite without subtracting those glyphs, the model fires phantom bounding boxes on the text and the recording dot. Cyrano computes a mask once per DVR layout at install (for example 4x4-std.mask and 5x5-std.mask) and caches the polygon coordinates by a layout_id key. On every frame, inference multiplies the composite by the cached mask in constant time. The mask is the artifact that lets composite-frame edge AI work at all; it does not exist in the greenfield-sensor edge AI story because the greenfield sensor has no overlay to subtract.

Yes, and intentionally. The HDMI composite is whatever the DVR decided the guard monitor should see: typically 1920x1080 at 15 to 30 fps, with each of 16 tiles landing around 480x270. That is lower than the native camera feed of, say, 2560x1440 at 20 fps. For the incident classes Cyrano targets (person in a zone, dwell threshold crossed, tailgate at a vestibule, package left too long) 480x270 per tile is enough. For applications that genuinely need full per-camera resolution (license plate read at 60 feet, facial recognition at distance) the side-channel form factor is the wrong tool and a greenfield-sensor system is the right one. The tradeoff is the point; it buys a 2-minute retrofit on hardware you already own.

On the Cyrano unit itself, at the property. The device holds the model weights, the filter stack, the layout cache, and the event queue on local storage. Frames do not leave the property during inference; only the already-compressed delivered events (thumbnail plus metadata) leave the device, and they go to the customer's WhatsApp or SMS endpoint. That is what lets the product honor the 'AI at the edge' privacy narrative on hardware that was never designed for edge AI.

Verkada and Rhombus are greenfield edge AI. They put a new processor inside a new camera and require ripping out the existing install. That works if the property is new construction or if the owner was going to replace cameras anyway; it costs $10,000 to $25,000 per property in hardware alone, plus labor. Side-channel edge AI skips the replacement and accepts the lower composite resolution in exchange. The two forms are not direct competitors; they serve buyers with different constraints. A Class A high-rise installing a new building chooses Verkada. A Class B/C multifamily operator with 16 DVRs already on the wall across 40 properties chooses side-channel.

It is not calibrated to the DVR model; it is calibrated to the layout. When the Cyrano unit sees a new composite layout on HDMI, it walks the frame edges to find tile boundaries, looks for the clock glyph region at known corner anchors, looks for the name-strip region at the tile header, and builds a polygon mask for all three categories. The mask is stored as a file keyed by a hash of the tile geometry plus corner anchors, with a human-readable alias such as 4x4-std or 5x5-std. After a guard switches between multiview and fullscreen a few times, the cache ends up with both 4x4-std and 1x1-std entries, which is the healthy shape. That per-layout calibration is what makes the retrofit device-agnostic across Hikvision, Dahua, Lorex, Swann, Uniview, Annke, Reolink, Night Owl, Amcrest, Q-See, and the long tail of rebranded DVRs shipped in the last decade.

Under 30 minutes per property in the field, and under two minutes for the physical wiring. You unplug the HDMI cable from the guard monitor, run it through the Cyrano box, run a second HDMI from the Cyrano pass-through to the monitor, and plug the unit's ethernet into the property network. The monitor keeps showing exactly what it was showing before. Zones and dwell thresholds are configured from the dashboard afterward. That is the whole physical install.

A WhatsApp or SMS message containing a tile thumbnail cropped from the composite (not the full multiview), a zone label, a dwell in seconds, a camera name, the layout_id that was active at detection, an end-to-end latency_ms from frame capture to message send, and an event class such as pre_action_zone_entry or loitering_dwell_exceeded. The responder can triage the message in under a minute without opening the dashboard. That last part is the half of edge AI that the greenfield-sensor narrative rarely discusses: what the human receives at the other end.