The only spec of an AI powered surveillance camera that matters is the 18 KB thumbnail on the on-call phone.

Because it is the only frame a human actually looks at. An AI surveillance camera produces one output that matters: a push notification on an on-call phone with one image that the operator glances at while walking, driving, or at dinner. If that image is a full-frame 4K screenshot of a 25-tile DVR multiview, the tile that fired is one twenty-fifth of the picture, unreadable on a phone lockscreen. If it is a tightly cropped 320x240 ~18 KB JPEG of just the tile that fired with the zone label printed, the operator can make a decision in the 2 seconds it takes to peek at the phone. Sensor megapixels upstream do not change that. The crop does.

Exactly three artifacts, ~240 KB total. One 320x240 JPEG thumbnail crop of the tile that fired, roughly 18 KB. One six-second clip of that same tile around the detection, roughly 220 KB. One JSON metadata object of about 612 bytes containing zone_id, dwell_seconds, wall_clock, class (person, vehicle, package), confidence, tile_index, layout_id, latency_ms, and event_class (for example pre_action_zone_entry or loitering_dwell_exceeded). Nothing else leaves: no full 1080p multiview frame, no raw video of the 24 other tiles, no continuous upload, no face embeddings, no plate strings. The event packet is the product boundary.

About 1.1 seconds in typical deployments. The budget breaks down roughly as 2 ms to decode the composite HDMI frame to RGB, 18 ms to run the one detection pass across all 25 tiles, 4 ms to apply per-tile zone rules and build the event packet, then the rest is network RTT to the cloud control plane and WhatsApp push delivery. The dominant cost is the last hop, not the model. If you are shopping AI surveillance cameras and someone quotes you 5 seconds latency, they are probably uploading full frames and running inference in a vendor cloud.

Three reasons. First, attention. On a phone lockscreen, a 320x240 crop of the firing tile fills the entire preview area; a 1080p multiview shrinks that tile to an unreadable 216x216 thumbnail. Second, bandwidth. Residential and small-business internet has upload caps; shipping the full 1080p frame per event adds up to tens of megabytes per day and gets rate-limited on cellular fallback. Third, privacy. The 24 tiles that did not fire may be showing a neighbor's hallway, a resident's door, or the inside of the mailroom. There is no reason to transmit those pixels off-property just because a different tile triggered. The crop keeps the event-of-interest and leaves the rest in the building.

zone_id (which named rectangle was crossed), dwell_seconds (how long the subject stayed in the zone before the event fired), wall_clock (the DVR's time), class (person, vehicle, package), confidence (the model score), tile_index (which of the 25 tiles), layout_id (which DVR multiview configuration was active, for example 5x5-std or 1x1-std after a fullscreen zoom), latency_ms (frame capture to message dispatch), and event_class (pre_action_zone_entry, loitering_dwell_exceeded, motion_post_curfew). These fields are what a responder uses to decide whether to ignore, talk down over the intercom, call, or escalate; and what an integrator uses to diagnose regressions. A system that delivers a picture with no metadata forces the human to re-interpret the scene every time, which is why most AI camera alert streams get muted within two weeks.

Existing. Cyrano is not a camera; it is a small edge device that plugs into the HDMI output of your existing DVR or NVR and runs inference on the composite multiview frame the DVR already draws for the guard monitor. That means the sensor you bought in 2018 (analog BNC, PoE IP, whatever) becomes an AI powered surveillance camera without any firmware change, credential setup, or network reconfiguration. One HDMI cable in, one HDMI cable out to the guard monitor, ethernet, power. Install takes under two minutes on a running DVR. The thumbnail that ends up on the on-call phone is cropped from the DVR tile of your existing camera.

The pricing shape is different because the output shape is different. Per-camera cloud products ship full camera frames (or heavily compressed streams) to the vendor's cloud continuously, pay for ingest, storage, GPU inference, and egress per camera, and render alerts with full-frame context. That model prices at $20 to $120 per camera per month. Cyrano processes the composite multiview locally, ships only the ~240 KB event packet off-property, and a single edge device covers up to 25 cameras off one HDMI input. At a 16-camera property on Cyrano's $200 per month per property price, the per-camera math is roughly $12.50. The delivered artifact on the phone is shaped to a phone, not to a cloud dashboard.

On a 16-camera multifamily property, a typical day produces 200 to 300 raw person-detections from residents, vendors, Amazon drivers, and passersby. A deployment-ready AI powered surveillance camera setup compresses that to 3 to 8 delivered alerts per day, a roughly 50-to-1 ratio. Below 25-to-1 and the channel gets muted in two weeks because the alert stream is noise. Above 200-to-1 and the system is dropping real incidents. Target a compression band of 40 to 100 raw events per delivered alert. The compression is done by zone rules, dwell thresholds, and event classes applied to the detection stream, not by a secondary model. The 3 to 8 that make it through are the events the phone shows a thumbnail for.

The adapter re-scopes. Every event payload carries a layout_id field; when the DVR operator double-clicks a tile on the guard monitor to zoom, the HDMI signal changes from the grid layout (for example 5x5-std) to a single-tile layout (1x1-std) at full resolution. The Cyrano unit detects the layout change, re-runs the zone-mask calibration for the new layout, and the thumbnail that gets delivered to the phone during fullscreen is now a 320x240 crop of the full-resolution single-camera frame. The layout_id in the JSON payload records which mode was active, so the integrator can tell from the event log whether the operator was zoomed in or not. The phone output stays the same size; only the source pixels changed.

Because responders already have WhatsApp open. The cost of an AI surveillance camera alert system is not the detection, it is the probability that the alert gets read within 30 seconds. A dedicated vendor app is a notification channel the on-call person has to remember to install, keep logged in, and not snooze. A WhatsApp thread with on-call staff is a channel they already check every few minutes. The ~240 KB event packet is formatted to render as a single push notification with an inline image preview, a caption that reads camera-name / zone-label / dwell-seconds / class, and a link to the six-second clip. The thumbnail renders in the push preview; the operator does not need to open the app to triage most events.