AI theft detection outside retail runs on a pre-action window, not a pose model.

Retail models classify shopper pose and gait: a hand going to a pocket, a coat held in front of a shelf, an item leaving the aisle without a POS scan. Outside retail, the actor is not a shopper and the scene is not an aisle. A person walking toward a transformer at 3 a.m. has no pose signature. A person standing near a mailroom package at 7 p.m. has no pose signature. What separates the ordinary resident from the thief is zone (are they in a place they should not be), dwell (how long have they been there), and time (is this within the window this zone is armed). Pose classification on its own fires on half of the residents walking their dog. The reliable signal for non-retail theft is person detection plus zone and dwell and time, not a pose model.

Watch footage of any non-retail theft and time the segment between the actor stopping in the target area and the actual act (cutting, lifting, prying). It is almost always between 30 and 180 seconds. That is the window where intervention still works. Inside it, a talkdown, a dispatch, or a loud speaker ping changes the outcome. Outside it, you are doing an insurance claim and reviewing footage. Detection that fires in minute 4 is a recording. Detection that fires in second 20 is a prevention. Non-retail AI theft detection has to be tuned around that window, which is why latency from tile to notification is a core spec, not a nice-to-have.

Cyrano does not pull the IP streams from individual cameras. It taps the DVR's HDMI multiview output, the same signal that drives the guard monitor. That composite video already has every camera on the recorder mosaiced into tiles, so one HDMI tap gives inference access to every feed at once. This works identically for analog cameras, IP cameras, rebranded Hikvision, Dahua, Lorex, Uniview, and any DVR or NVR with an HDMI port. No camera firmware involvement, no ONVIF negotiation, no credentials to coordinate. Physical install on a running DVR is under 2 minutes: HDMI in from the DVR, HDMI out to the monitor, network cable, power.

Yes, and this is the part most vendors skip. Every DVR overlays a clock, a per-tile camera name strip, and a channel bug. To a detection model, those glyphs are visual noise that can trigger on anything from a bounding box on a camera name to a false positive on the clock's colon. Cyrano's install step labels each tile and masks the fixed-position overlays before inference so the model only sees the scene. This happens once per DVR layout, not once per frame.

Up to 25 tiles per unit off a single DVR multiview. If the DVR is set to a 4x4 grid the unit runs inference across 16 tiles in parallel. If it is a 5x5 it runs across 25. If the operator switches the DVR to fullscreen on a specific camera, the unit re-scopes to that single camera at full resolution, so during an active incident you actually get higher per-tile accuracy, not lower. A property with more than 25 cameras typically runs one Cyrano unit per DVR.

Because the channel that gets read is the channel already open on the phone. Property staff already use WhatsApp for maintenance, move-outs, and vendor coordination. Adding one more thread is a free change. Adding another app with another login is how alerts end up unread. Dedicated monitoring portals start around $250 per camera per month because they assume a human watching a feed. One WhatsApp thread per property, with a tile thumbnail and a timestamp on every alert, matches how Class B and C multifamily, construction, and small commercial teams already communicate. It is a deliberate product choice, not a shortcut.

The categories where pre-action detection has the most leverage: package theft from mailrooms and lobbies, cable and copper theft at transformer pads and conduit runs, HVAC theft (copper line sets, condenser units), cargo theft at loading docks and trailer yards, parking-lot theft of vehicles and catalytic converters, and jobsite theft of tools, wire, and material. Every one of these has a pre-action window where the actor enters a defined zone, pauses, assesses, and then acts. Retail shoplifting detection models do not cover any of these, because none of them involve a register, an aisle, or a shopper gait.

Yes, because the entire pipeline is scoped. A raccoon, a trash bag blowing across the lot, a delivery driver turning around, or a shadow at sunset triggers a motion sensor and mutes the alert channel within a week. Person detection plus a drawn zone plus a dwell threshold plus a time window is three filter layers compared to one. In production deployments, a typical 16-camera property generates a handful of read-able alerts per day rather than dozens per night. That is the difference between an alert channel that stays in service and one that gets muted.