IREX
EN

Platform architecture

How video travels from a camera to an operator, the three ways a site can connect, and the open-source foundation the platform runs on. This page is the conceptual map; the component-level detail is in Microservice architecture.

The path from camera to operator

IREX platform data flow, from camera to operator SOURCES CONNECTIVITY PROCESSING AND STORAGE ACCESS IP cameras ONVIF, RTSP Sensors and devices on user-defined map layers Uploaded media files added for analysis Direct connection fiber or broadband, no on-site server Edge server local AI and storage IREX Private Cloud Kubernetes, 100% open-source stack Ingest and archive video chunks, snapshots, retention Analytics pipelines CPU (Synet) and GPU (NVIDIA CUDA) Event bus detections, matches, alarms Search index clustering for sub-second search Operator workspace map, video, search Alarm monitors real-time alerts Sover messenger field alerts, Ask IREX API and Superset integrations, dashboards Every request above passes role-based access control, the Case ID gate, and the audit log
Video and events move left to right. Access control and logging apply at every point where an operator, an integration, or an agent touches the platform.

Deployment models

Connectivity conditions differ site by site, so the platform supports three models. The commercial model is independent of the connectivity model: any of the three can be licensed as a subscription or as a perpetual license.

Model Where AI runs Suited to
Direct connection In the private cloud. No servers on site. Sites on fiber or high-availability broadband.
High-availability edge server On redundant edge servers at the site, over wired or wireless links. Bandwidth-constrained or unreliable connections where downtime is not acceptable.
Cost-effective edge server On a single edge server at the site. Smaller sites with limited connectivity.

All three provide centralized management, high availability, access logs, and web access from any device. See Edge servers and Deploy instance.

How the platform scales

Four dimensions scale independently and linearly: camera count, user count, retention period, and the number of active AI modules. Doubling any one of them is an operational change, which means adding nodes, not an architectural change.

Dimension Proven in production Architectural ceiling
Cameras on one instance 100,000+, all running real-time analytics 1,000,000
Concurrent users Multi-agency command centers 100,000
Instant-search retention 30 days 1 year
Active AI modules per channel See Module compatibility Limited by node capacity, not by design

Storage grows by adding Ceph nodes; the descriptor index grows with the data plane; analytics workers are stateless and scale horizontally. See Scale instance and System requirements.

Resilience

Public safety deployments run continuously for years, so component failure is treated as normal operation rather than as an incident:

  • No single point of failure. Any server, disk, switch, edge node, or analytics worker can fail or be replaced under load without downtime visible to an operator.
  • Triple data replication in Ceph distributed storage, with automatic failover across control-plane and data-plane services.
  • Kubernetes-native rolling restarts for upgrades, so a version change does not require a maintenance window. See Update instance.
  • Redundant edge servers in the high-availability deployment model, so a site keeps recording and analyzing when its uplink drops.

The open-source foundation

The infrastructure layer is entirely open-source, with no third-party proprietary code. This is what makes the stack inspectable and what removes any dependency on a vendor's hosted service:

Component Role in the platform
Kubernetes Orchestrates every service; restarts failed containers and migrates them off failed hosts.
Ceph Distributed storage for the video archive, with triple replication.
Cassandra Stores raw video data and full event records at petabyte scale.
Apache Ignite In-memory store for event metadata, which is what makes search return in under a second.
Kafka Message broker between all services; carries events, matches, and alarms.
PostgreSQL Relational store for configuration, users, permissions, and resources.
Elasticsearch and Apache Spark Indexing and batch analytics behind reporting and dashboards.

A Leaf-Spine network architecture keeps latency predictable across the cluster. The whole stack is hosted by the IREX Private Cloud, IrexAI's bare-metal-to-Kubernetes platform, which also hosts Sover.

Service-by-service detail, including the microservice names and their inputs and outputs, is in Microservice architecture. What this foundation means for ownership and auditability is in Data and technology sovereignty.