AI model · Face recognition
Engineered for accuracy. Built for real-world impact.
Our facial recognition blends deep learning, biometric science, and edge AI to deliver real-time identification with 99.82% accuracy—even in demanding environments.
Overview
See it in context
AVA-X face recognition in a real deployment scenario—play when you are ready.
Principles
Philosophy of precision
Our technology is built on principles that carry through from research to deployment.
Swiss-made AI
Engineered in Switzerland with precision, reliability and performance at the core.
Ethical AI
Privacy, security, and transparency as fundamentals—not afterthoughts.
Real-world ready
Built for practical deployment in challenging environments, not only lab conditions.
Architecture
Core technology pillars
Four modular components that work together for performance, accuracy, and security.
- Multi-angle detection, even with occlusions
- Real-time 1:N face matching with millisecond latency
- Based on proprietary Safeza AVA-X deep learning models
- Optimized for NVIDIA Jetson edge devices and HPE servers
Face detection & landmarking
Multi-angle detection with occlusions; landmarks under varied lighting and camera quality.
Face embedding & vectorization
Encrypted biometric signatures from AVA-X models—robust to age, glasses, and facial hair.
Face matching engine
Real-time 1:N matching from edge lists to large databases—Jetson to data centre scale.
Smart alerting & API integration
Blacklist, whitelist, anomaly logic; SMS, Telegram, Threema, HTTPS, Kafka—and VMS workflows.
Trust
Compliant with leading security and privacy expectations
POPIA, GDPR, and Swiss data protection—engineered for auditable, responsible deployment.
ISO 27001
GDPR
CCPA
SOC2
Swiss engineering & precision
Engineered in Switzerland with reliability and performance at the core.
Industry-leading accuracy
99.82% accuracy that exceeds typical benchmarks—trusted in production.
Data privacy & compliance
Aligned with GDPR, POPIA, and Swiss data protection expectations.
Benchmarks
Outperforming global standards in face recognition accuracy
Validated against industry references—built for production scenarios, not slides alone.
Benchmark comparison: LFW dataset accuracy
+0.19% higher than the reference figure shown—fewer false positives and negatives where it matters.
Accuracy on LFW
99.82%
Reference-class performance
Processing time per face
< 0.5s
Low latency for live use
Simultaneous face processing
10,000+
Crowded scenes and dense feeds
Infrastructure
Powerful AI where you need it—on-premise or in the field
From edge devices to high-performance clusters: flexible deployment without compromising control.
| Device | NVIDIA Jetson Nano | Minimum: 128-core NVIDIA Maxwell GPU, 4GB RAM, 5W power consumption |
| Device | NVIDIA Jetson NX | Minimum: 384-core NVIDIA Volta GPU, 8GB RAM, 15W power consumption |
| Device | NVIDIA Jetson AGX | Minimum: 512-core NVIDIA Volta GPU, 32GB RAM, 30W power consumption |
| Device | NVIDIA Jetson Orin | Minimum: 1024-core NVIDIA Ampere GPU, 64GB RAM, 60W power consumption |
| Server | HPE ProLiant Servers | Minimum: Dual socket support, up to 3TB RAM, enterprise-grade reliability |
| GPU | NVIDIA Tesla L40 GPUs | Minimum: 48GB GDDR6, 142 RT cores, 568 Tensor cores, 300W TDP |
| GPU | NVIDIA RTX A6000 | Minimum: 48GB GDDR6, 84 RT cores, 336 Tensor cores, 300W TDP |
| Processor | Intel Xeon Processors | Minimum: Up to 56 cores, 3.5GHz base clock, support for AVX-512 |
| Feature | Bandwidth efficient | Compressed video streams, adaptive bitrate, <1Mbps per camera |
| Feature | Parallel processing | Multi-threaded architecture, GPU acceleration, async I/O |
| Feature | Air-gap compatibility | No external network required, local processing, isolated deployment |
| Feature | Offline operation | On-device AI inference, local database, zero cloud dependency |
| Feature | Encrypted communication | TLS 1.3, AES-256 encryption, certificate-based authentication |
| Feature | Secure boot | UEFI secure boot, hardware root of trust, signed firmware |
| Feature | Multi-factor authentication | TOTP, hardware tokens, biometric verification support |
| Feature | Role-based access | RBAC implementation, granular permissions, audit logging |
Infrastructure flexibility
From portable edge units to on-premise command centres—performance tuned to how you operate.
Field deployment
Self-contained units for mobile operations, checkpoints, and remote locations.
Command centre
High-capacity processing for centralised monitoring and large-scale operations.
Field performance
Accuracy in the wild
Laboratory benchmarks matter—but production is where discipline shows. Built for unpredictable environments.
Crowds & high density
10,000+ faces in dense urban environments
Low-light conditions
Precision in challenging lighting
Wide range of cameras
From RTSP to USB
Occlusion & disguises
Partial faces, aging, accessories
Designed for challenging conditions
Extreme angles and partial faces
Accurate detection even with challenging angles
Motion blur from moving targets
Fast-moving subjects with usable results
Cross-camera matching
Tracking across multiple feeds
Legacy CCTV integration
Works with existing infrastructure
Multi-facial feature recognition
Advanced multi-point analysis
Resilience to spoofing
Protection against fraudulent attempts
Responsibility
Built with ethics. Designed for compliance.
Privacy and governance are part of the architecture—not a bolt-on checklist.
GDPR compliance
European data protection requirements
Enterprise-grade encryption
Biometric data protected in transit and at rest
Minimal data storage
Embedding vectors—no unnecessary retention
Role-based access
Strict control over who can use the system
Next step
From facial features to actionable intelligence
See how Swiss-engineered AVA-X fits your access, live, and investigation workflows—with accuracy you can stand behind.
