Platform
Canopy sits downstream of your existing infrastructure. No rip-and-replace. No cloud egress of video. Edge inference runs locally; only structured intelligence leaves the site.
The platform is designed as a strict one-way data flow. Video enters at the edge; structured intelligence exits to your dashboards, APIs, and operators. Nothing in that chain requires raw footage to travel.
RTSP / ONVIF / HTTP streams from any IP camera, NVR, or DVR. Canopy Edge Gateway connects to the existing CCTV control room with no rewiring.
Models run on-device. Detections, counts, trajectories, and events are computed locally. Frames are discarded after inference — never buffered to disk.
Structured signals are batched, encrypted, and forwarded to the Canopy Intelligence Cloud. The payload contains counts and events — not pixels.
Dashboards, REST API, webhook delivery, and the conversational AI layer. Operators query intelligence in plain language or consume it in their BI stack.
Canopy Edge runs on commodity hardware — a compact x86 or ARM appliance that connects to the camera LAN. No GPU cloud required. Inference latency is sub-100 ms per frame at full 1080p.
Because the model runs on-site, raw video never crosses a network boundary. The appliance holds a rolling inference buffer of two seconds for event confirmation; nothing is persisted beyond that window.
streams:
active: 24
degraded: 0
offline: 0
inference:
model: canopy-vision-v3.2.1
backend: tensorrt
latency_p95: 82ms
fps_avg: 28.4
privacy:
frame_buffer: 2s (ring)
disk_persist: disabled
pii_filter: enabled
uptime: 99.97%General-purpose object detection is the wrong starting point for commercial intelligence. Canopy trains vertical-specific models that understand domain context — a mall footfall model knows what a zone boundary is; a construction safety model knows what a hard-hat exclusion zone is.
Multi-label detection of people, vehicles, and objects with spatial positioning relative to pre-defined zones. Calibrated for the camera angles and densities common in commercial property.
Re-identification across camera views within a session — without storing any biometric signature. Trajectories are tokenised per session; no cross-session linking is possible.
Head-pose and body-orientation inference for media attention measurement. Panel viewability is computed per frame; no facial geometry is extracted or stored.
Helmet, vest, and harness compliance per zone on construction sites. Alerts trigger on first non-compliance frame; no footage clip is generated for the alert.
Aggregate headcount and vehicle count per zone with temporal bucketing. Outputs a time-series of counts; no individual identifiers are present in the output.
Queue formation, crowd density spikes, loitering, direction-of-flow reversal. Event payloads contain zone, timestamp, and magnitude — nothing else.
Once edge inference produces a structured event, the pipeline handles normalisation, aggregation, and delivery. The pipeline runs on Canopy-managed infrastructure within the operator's chosen jurisdiction — UAE or EU.
Data is partitioned by site, by vertical, and by time bucket. Downstream consumers — dashboards, APIs, BI connectors — always see aggregated series, never the raw event stream.
{
"site_id": "mall-al-nakheel-01",
"zone": "west-atrium",
"ts": "2026-06-07T14:32:00Z",
"bucket": "5m",
"signals": {
"headcount": 312,
"entries": 47,
"exits": 39,
"dwell_avg_s": 312,
"occupancy_pct": 68
},
"pii": false,
"footage_retained": false
}Edge gateway forwards encrypted structured events over mTLS. No video payload.
Events are schema-validated, clock-synced, and tagged with site and zone metadata.
Raw events are bucketed into 1-min, 5-min, 1-hour, and 1-day series. Raw event log is deleted after 24 hours.
Aggregated series are retained for 36 months by default. Retention window is configurable per operator contract.
REST API, WebSocket push, and dashboard reads all resolve against the aggregated store. No access to the raw event log.
Dashboards answer the questions you already know to ask. The conversational layer handles everything else. Operators query their full intelligence estate in plain language — Canopy routes the query, retrieves the relevant signals, and responds with a sourced answer.
Every response is sourced to a specific site, zone, and time range. The model cannot fabricate metrics.
Queries are scoped to the sites and metrics the authenticated user is permitted to see — enforced at the data layer, not in the prompt.
Operators can configure threshold alerts conversationally. No dashboard UI required for common monitoring tasks.
Every conversational query can be exported as a REST API call. The intelligence layer and the API share the same data access layer.
Canopy is not a destination dashboard. Intelligence should live wherever operators already make decisions — BI tools, property management systems, operations centres.
Authenticated, versioned endpoints for all aggregated series. OpenAPI spec available on request.
Event-driven delivery for threshold alerts and anomaly detections. Configurable per site and zone.
Native connectors for live data refresh. Operators embed Canopy metrics alongside their existing property KPIs.
Structured integration with Yardi, MRI, and custom PMS platforms via API. Footfall and occupancy sync on configurable cadence.
OPC-UA bridge for building management systems. Occupancy signals feed HVAC and access-control decisions in real time.
Syslog and CEF export for security operations centres. Behavioural anomalies route to your existing incident workflow.
Bring a site plan or a live RTSP URL. We'll show you what Canopy would surface within the first hour of connection.
Book a technical demo