Electricity theft detection, for the person who owns the panel, not the utility that owns the meter.

Utilities detect electricity theft by running machine learning over smart meter telemetry. Random forests, gradient boosting, and deep CNNs flag accounts whose consumption drops without a matching drop in downstream load. That is the academic SERP for this keyword, and it is solving the utility's problem: finding customers who under-report. A property owner has a different problem. The anomaly is already visible on the monthly bill or the submeter report. What is missing is who plugged into the exterior pedestal outside unit 312, who opened the service-panel room at 02:14, who jumpered the submeter in the electrical closet on the 3rd floor. That is not a telemetry problem. It is an identification problem at a physical point, which is what a camera zone event solves.

Five patterns account for almost every case. First, exterior 30A and 50A pedestals (pool equipment power, RV hookups, courtyard outlets) used by neighbors running a cord across a fence for a space heater, a hot tub pump, or a grow tent. Second, 240V laundry outlets tapped after hours for EV charging or a window AC fed into a non-resident unit. Third, bypassed or jumpered submeters in the electrical closet, where a contractor or resident clips around the submeter so load reads zero while a unit or sub-space still consumes. Fourth, opened service-panel rooms where a tap is pulled off an unmetered bus. Fifth, common-area outlets on house panels used for heavy loads (mining rigs, welders, electric smokers) that never appear on any tenant bill. The common factor: each happens at a known physical point the owner already has a camera looking at or can point one at.

Cyrano is an edge AI device that taps the DVR or NVR's HDMI multiview output. That output is the same composite the guard monitor already shows, with up to 25 camera tiles mosaiced in one frame. Cyrano pulls the HDMI signal, runs per-tile person detection and zone classification locally, and emits a zone entry event when someone enters a drawn polygon and dwells longer than the configured threshold. For electricity theft, the zones are: meter_bank_access_door, submeter_closet_3rd_floor, exterior_pedestal_bay_17, service_panel_room, house_panel_common_area_outlet, laundry_240v_outlet_wall. The detection runs on-site, the event lands on the property's WhatsApp thread in seconds, and the payload pairs one-to-one with the kWh anomaly on the same submeter.

No. The capture point is the DVR's HDMI output driving the guard monitor. That composite already has every camera on the recorder. One HDMI tap gives inference access to every feed at once. Works on analog cameras, IP cameras, any DVR brand including Hikvision, Dahua, Lorex, Amcrest, Reolink, Uniview, Swann, Night Owl, Q-See, and most rebrands. No camera firmware involvement, no ONVIF negotiation, no credentials to coordinate. Physical install on a running DVR is under 2 minutes: HDMI in, HDMI out to the monitor, network, power. One unit handles up to 25 tiles.

Smart meter models detect non-technical losses by comparing consumption patterns against neighbors, against historical baselines, or against feeder-level energy balance. They work well for utilities catching end customers who under-report. They fail for property owners for three reasons. First, the tap is often upstream of the submeter (exterior pedestal, house panel, service entrance) so the submeter reading is already correct and no anomaly appears in that meter's data. Second, the theft signature is dwarfed by legitimate load variation in multifamily and commercial buildings. Third, and most importantly, even when the ML correctly flags an anomaly, it produces a number on a dashboard, not a person at a specific second. The dashboard tells you something happened. The camera zone event tells you who. Without the who, the submeter anomaly is a billing dispute with no defendant.

event.class (pre_action_zone_entry), property identifier, tile.label (camera name), zone name (for electricity theft, typically meter_bank_access_door, submeter_closet_X, exterior_pedestal_N, service_panel_room, laundry_240v_outlet, house_panel_outlet), zone.armed_window, dwell_seconds (integer), timestamp (ISO 8601 with timezone offset, to the second), a tile thumbnail URL (480x270 crop centered on the event), overlay_mask identifier (which DVR overlays were masked before inference), delivery channel (WhatsApp group or SMS fallback), and latency_ms (capture to delivery). The payload is deliberately minimal and deliberately machine-readable. It is the exhibit that pairs with the submeter's monthly kWh delta on the recovery ticket.

Five documents. Stage 1, submeter reading showing the kWh anomaly on a specific circuit or unit, pulled from the building's submetering system (InterSolv, Nextek, Metron-Farnier, eWater, or the owner's chosen vendor). Stage 2, the zone configuration showing which camera polygon was drawn around which physical access point, with what dwell threshold and armed window. Stage 3, the Cyrano zone entry event itself (tile thumbnail, zone name, dwell count, ISO timestamp). Stage 4, a correlation document showing the event timestamp inside the kWh anomaly window. Stage 5, for recovery, a signed application log (for exterior pedestals, a ledger of who has access cards, for service panel rooms, an access-control log). That is the file that converts an unrecoverable billing loss into a collectable or prosecutable case.

The pattern holds from 20-unit single buildings up to 400-unit garden-style portfolios. Small properties actually see the clearest signal because the population of legitimate users at any given physical point (the service panel room, the pool equipment pedestal, the laundry 240V outlet) is small, so every event is meaningful. A 30-unit property with one exterior RV pedestal and one service-panel room is an ideal deployment: two zones, one camera unit, and the monthly submeter anomaly is already known. Large sites need more zones and better camera placement but the economics are the same: a $450 one-time hardware unit plus $200/month handles up to 25 camera tiles, versus a $3,000 to $5,000 monthly guard that cannot be in the electrical closet at 02:14 anyway.

If the theft is upstream of the building main (a clip on the utility service drop, a bypass at the meter can before it reaches your panel), then yes, the camera event is the evidence the utility will actually accept. Utilities run their own ML on consumption data and will flag the anomaly. When they open an investigation, the document they need is a timestamped record of a person at the meter can or service entrance. That is exactly the zone entry event. Utilities in Georgia, Louisiana, and Texas routinely accept this kind of evidence as part of the investigation file, and some cooperatives have standing requests for any property with repeat theft to add a camera zone on the meter can for this reason.

If the camera already there is IR-capable (most current PoE cameras are), yes. Cyrano works off whatever image the DVR is recording. For unlit spaces, the fix is a low-lux or IR-illuminated camera on the electrical closet, submeter cabinet, or service panel room door. The detection itself is image-based and does not care about lighting source, only whether a human is visible in the frame. For exterior pedestals, standard IR flood coverage works. For an interior closet, adding a small IR LED bar if the existing camera is not IR-capable is typically a $40 fix.