Performance and scale

How many cameras can the platform handle?

A single IREX instance is proven in production at 100,000+ cameras, all running real-time AI-powered video analytics with 30-day instant search. Across all deployments, IREX manages 250,000+ cameras in 8 countries — the United States, the United Kingdom, Ireland, Peru, Israel, Bosnia and Herzegovina, Mongolia, and Kazakhstan. The architecture scales horizontally and linearly to 1,000,000 cameras, 100,000 users, and a one-year retention period without architectural change: adding cameras means adding Kubernetes and Ceph nodes, deeper archives mean adding storage nodes, and more users mean adding balancer nodes. Sizing for a specific deployment — server count, CPU/GPU mix, storage — is calculated per project by IREX engineering. See Scale an instance in the user guide.

What operational results has the platform delivered in production?

MetroCCTV, an IREX partner in the UK and Ireland, reports the following operational outcomes from production use:

• 1,000+ cameras per operator in the control room — one operator running IREX covers a camera count that previously required a wall of analysts.
• Under 1 second to search video data across petabyte-scale archives.
• 10x or greater reduction in average response time, compared with manual video surveillance and citizen-report (911-style) workflows.
• 100x or greater reduction in average investigation time, compared with manual playback review driven by a motion detector.

These are partner-reported operational figures from live deployments, not laboratory benchmarks; results depend on camera infrastructure, staffing model, and operating procedures.

Has IREX facial recognition been independently benchmarked?

Yes. IREX submitted its face recognition algorithm to the NIST Face Recognition Vendor Test (FRVT), the US government's independent benchmark. In the results NIST published in March 2021, the IREX algorithm ranked #1 in accuracy among all US companies, and in the top 10 globally out of 268 algorithms submitted worldwide, on the Border and Kiosk datasets of the FRVT 1:N Identification track. Those two datasets are the closest FRVT sets to street-camera conditions — less constrained in viewing angle, lighting, and resolution than the other FRVT datasets — which is why IREX reports them. This is a dated result from the March 2021 submission and does not represent a current or continuing ranking. The engine behind it is trained on a proprietary dataset of 40 million images optimized for real-world CCTV conditions, and it adds a layer that aggregates templates across multiple video frames, which a still-image benchmark such as FRVT does not measure.

How does IREX address accuracy and bias in facial recognition?

IREX treats bias awareness as one of the six pillars of its Transparency by Design framework, starting from a stated fact: all face recognition systems, including IREX's, show varying accuracy across demographic groups of gender, age, and skin color. The company's response has four parts:

• Balanced training data and algorithm work. IREX data scientists work continuously to reduce demographic bias by balancing training datasets and optimizing algorithms. The face recognition engine is trained on a proprietary 40-million-image dataset optimized for real-world CCTV conditions.
• Independent evaluation. IREX participates in the NIST Face Recognition Vendor Test, which provides third-party evaluation of accuracy and demographic effects.
• Own testing. IREX has conducted a company study to estimate bias in its suspect-recognition technology across demographic groups (published July 2024).
• Proactive disclosure. Where systematic bias exists, IREX policy is to disclose its existence and estimated impact to the customer rather than leave the customer to discover it.

Operationally, deployments are designed so accuracy questions can be audited: every high-risk action is bound to a Case ID and a non-erasable audit log, and detection accuracy trends, camera uptime, alert frequency, false-positive rates, and processing latency can all be analyzed per deployment in the platform's Apache Superset dashboards. No detection system is error-free, so IREX does not publish blanket accuracy percentages or promise guaranteed detection; where accuracy is claimed, the basis for it is stated.

Can multiple agencies share one deployment?

Yes. IREX has a carrier-grade, multi-tenant architecture with hierarchical role-based access control, which is what allows several agencies or jurisdictions to work on a single deployment without seeing each other's data. Users belong to a hierarchical catalog of user groups, each granted access to specific resources — cameras, watchlists, locations, venues, alarm monitors, and external devices — and a user can belong to several groups with a different role in each, chosen from six roles: Viewer, Member, Operator, Group Manager, Camera Manager, and System Manager. Platform features, databases, match alerts, and search results are visible only to users whose role permits them, which implements the access-control principles of CJIS and GDPR. On that basis, agencies can share cameras and evidence across jurisdictions and coordinate through the Sover end-to-end encrypted messenger, while every search and export remains bound to a Case ID and a non-erasable audit log. See Manage users and user groups.

What uptime and availability can we expect?

The platform is designed to survive component failure rather than to depend on any single machine. Storage is triple-replicated and failover between control- and data-plane services is automatic, so the system keeps running when individual servers or drives fail; Kubernetes handles self-recovery, load balancing, and migration of workloads off a failed node; and infrastructure components support rolling upgrades. For IREX-hosted service, the standard IREX SaaS Agreement commits IREX to commercially reasonable efforts to keep the hosted software available 99.5% or more of the time in any calendar month, backed by service credits if that standard is missed. The standard excludes notified planned maintenance, force majeure, customer-caused outages, and internet or ISP failures outside IREX's control, and it does not apply to features identified as beta. Customers who deploy on their own private cloud or on-premises hardware operate the instance themselves and set their own availability targets, with the same high-availability design underneath.