Situational intelligence built around the cameras you already have.
Phoenix All Eyes adds an event-first intelligence layer over existing cameras, NVRs, access control and GPS units: detection, correlation, search, prioritization and a traceable response workflow for authorized operators. It does not take coercive decisions and does not act outside the rules and permissions you define.
Confidential and adaptive by design: the same core, configured per operation.
Deployments are specific to each client’s infrastructure, risk and rules. What is shown publicly is architecture, not client sites.
Logistics networks
Yards, docks, cross-docks and fleets: dwell time, unauthorized access, unit ↔ camera correlation, incident evidence packs.
Public security
Municipal camera networks and C4/C5 rooms: prioritized events, false-alarm measurement, traceable response and an audit chain per intervention.
Private security
Corporate sites, retail and industrial plants that already own cameras and NVRs: an intelligence layer above them, not a replacement.
Remote and edge sites
Substations, ranches, towers, borders of operation: edge node on site, events and clips only — never continuous raw video over the link.
Seven levels, escalated by cost. Not every frame goes to a large model.
Ingest
RTSP/ONVIF from existing cameras and NVRs. Sub-stream for 24/7 analysis; main stream only on escalation. Cheap, always on.
LEVEL 0Perception
Light detector + tracker on every frame. Deterministic. Entities stay anonymous (person_42, vehicle_18) unless identity is explicitly authorized.
LEVEL 1Event engine
Rules turn tracks into events: zone entry, loitering, line crossing — with debounce, idempotent IDs and per-camera heartbeats so silence is never mistaken for calm.
LEVEL 2Selective escalation
Only the relevant clip window (T−30s to T+60s) goes to deeper analysis, gated by severity — not by volume. Low GPU, low bandwidth, low storage.
LEVEL 3Reasoning
Anomaly defined in plain text per site and zone; no model retraining per deployment. Three separate scores: anomaly, security risk, model confidence.
LEVEL 4Entity graph + GPS
Person ↔ vehicle ↔ place ↔ unit ↔ incident. GPS from field units and legacy trackers correlates with camera events (backward compatibility with existing GPS systems in progress).
LEVEL 5Operator review
Low confidence produces human review, never automatic escalation. Every alert keeps the original event, the inference, the human intervention and the outcome.
LEVEL 6Why this is ours — detectors and trackers are replaceable. The event schema, the rule engine, the correlation layer and the operator feedback loop that accumulates a proprietary dataset are not. That is where the moat is, the same way it was for the companies that defined this category.
What must be proven before we sell site operations.
A controlled demo on a cloud GPU is scheduled for October 2026. Multi-site promises come after a measured multi-camera beta — not before.
Multi-camera performance
Stable at the agreed resolution and FPS, 8–20 streams, one site.
False positives and negatives
Measured per priority event, against human labels, with a fixed adversarial sample set before the run.
Latency and recovery
End-to-end latency and behaviour under network loss: no lost events, no duplicated events on restart.
Real operator workflow
Tested with a real operator, plus retention, privacy, permissions and signage reviewed with applicable counsel.
Assessment first. Pilot second. Operations only after the gates.
Cameras, NVRs, network, GPS units, risks and workflow. Deliverable: integration map, data scope and a test plan. No hardware purchase.
8–12 weeks, measured detection and operator workflow. Edge hardware is quoted after the assessment: streams, resolution, retention and redundancy decide it.
Recurring operation, implementation and site node hardware quoted separately. Multi-site enterprise only after site surveys. Institutional deployments run on infrastructure inside the client’s perimeter.
Start with a site assessment.
Tell us what cameras, NVRs, network and units exist, and what the operation needs to see. We answer with an integration map and a test plan, not a promise.