A theft detection system has seven components. On your property, five of them are already installed.

In order from the optical end to the operator: (1) the camera sensor and lens, (2) the coax or ethernet transport, (3) the DVR or NVR that composites feeds and records to disk, (4) the signal-access layer that exposes video to a detection process, (5) the detection and classification model itself, (6) the scope filters that gate which detections become alerts (zone, dwell, armed time window), and (7) the egress channel that carries alerts to a reader. On any property with CCTV that is more than three years old, components 1 through 3 and component 7 are already installed and running. The actual work of adding a theft detection system is confined to components 4, 5, and 6.

Because every 'all-in-one' theft detection system on the market sells you the five components you already have bundled in with the two you do not. That is where the $50,000 to $100,000 per-property rip-and-replace quotes come from. The bundle math only works in vendor spreadsheets. If you separate the layers, the question 'what does it cost to add theft detection to this property' becomes a question about components 4, 5, and 6 only. Those three components can be delivered as a single $450 device and a $200 per month software subscription. The rest of the stack does not get touched.

By tapping the DVR's HDMI multiview output, which is the same composite signal that already drives the guard monitor. That signal carries every camera on the recorder mosaiced into tiles, so a single HDMI cable gives the inference pipeline access to every feed at once. This avoids ONVIF negotiation, per-camera credential management, firmware compatibility checks, and multicast network configuration, because all of that work has already been done by the DVR when it composed the multiview. The physical install is HDMI in from the DVR, HDMI out to the guard monitor, network cable, power. Under two minutes on a running system.

Every DVR burns its own on-screen graphics into the multiview output: a clock in one corner, a channel bug in another, and a per-tile camera name strip at the top or bottom of each tile. Those glyphs are visible in the pixels the inference model receives. Without special handling, a bounding-box model can misclassify a camera name text strip as an object, or a clock colon flicker as motion. Cyrano computes an overlay mask once per DVR layout_id (for example 4x4-std or 5x5-std), caches it, and applies it before every inference frame so the model never sees those fixed pixel regions. The mask is a lookup by layout_id, not a per-frame computation.

A delivered alert carries a tile thumbnail cropped from the DVR multiview, the polygon zone label that was crossed, the dwell seconds that were counted before the alert fired, a timestamp, the camera name, the DVR's layout_id (4x4-std, 5x5-std, etc., used to apply the overlay mask), and the end-to-end latency in milliseconds from frame capture to message send. The event class for a pre-action pattern is pre_action_zone_entry. Those fields are what make an alert actionable: a dispatcher opening the WhatsApp message can verify the scene, read the dwell count, and decide to talk down, dispatch, or log.

Up to 25 tiles off a single DVR HDMI multiview. If the DVR is set to a 4x4 grid the unit runs inference across 16 tiles in parallel; if it is set to a 5x5 it runs across 25. When an operator switches the DVR to fullscreen on a specific camera during an active incident, the unit re-scopes to that single camera at full resolution, so per-tile accuracy goes up during the moments that matter. A property with more than 25 cameras typically runs one unit per DVR.

No. All detection and classification happens on the Cyrano device, which sits in the same rack as the DVR. Only the alert payload (a small thumbnail, the zone label, dwell seconds, timestamp) leaves the building, and only when a zone-verified dwell-threshold-cleared in-window event fires. Raw video is never uploaded. This keeps bandwidth costs zero for the detection layer and keeps the system compatible with tenant-privacy expectations in multifamily.

Package theft at mailroom doors and lobby shelves, cable and copper theft at transformer pads and conduit runs, HVAC theft at condenser cages and line-set chases, cargo theft at loading-dock aprons and trailer yards, parking-lot theft including vehicle entry and catalytic converter removal, and jobsite theft from conex boxes and staging areas. Each of these shares a structural pattern the filter stack maps onto cleanly: a defined target zone, a pre-action pause while the actor positions, and a time window during which legitimate presence is near zero.

Hardware install is under 30 minutes, counting cable routing. Zone configuration (drawing polygons on each camera tile and setting the armed time windows) typically takes 15 to 45 minutes per property depending on camera count and how many zones the operator wants. A 16-camera multifamily property with 8 zones is usually alert-ready inside one afternoon. The tagline is not marketing filler: 'live in one afternoon, not one quarter' is a deliberate contrast with the 8-to-12-week timeline of a camera-replacement project.

Cyrano is $450 one-time for the hardware plus $200 per month for the software starting month two. For a 16-camera multifamily property that is $450 + ($200 * 12) = $2,850 in year one and $2,400 per year thereafter. A full-replacement smart-camera deployment on the same property typically runs $50,000 to $100,000 up front plus $250 per camera per month in cloud subscription fees. On a 16-camera site the bundled model crosses $100,000 within three years. A single on-site guard runs $3,000 to $5,000 per month, which is more per month than Cyrano costs for two years.