About the client
QuantumEye is an AI company developing proactive, ethical surveillance technology to support retail security operations. The organization aims to transform traditional CCTV systems into intelligent, automated solutions for preventing theft. Their mission focuses on protecting businesses and communities while maintaining strict adherence to customer privacy and operational integrity.
Business context
UK retailers are facing an unprecedented surge in retail crime, from organized shoplifting groups to rapid grab-and-run incidents and highly frequent repeat offenders. Losses are escalating quickly, with retailers now losing more than £2.2B annually to theft.
Legacy CCTV systems are fundamentally reactive. They require continuous manual monitoring and typically provide evidence only after an event has occurred. This operational gap leads to:
- missed detections and rising shrinkage
- staff overstretch and increased safety risks on the shop floor
- inability to stop repeat offenders in real time
QuantumEye needed to bridge this gap by transforming existing CCTV infrastructure into an AI-powered prevention system that detects suspicious behavior the moment it occurs, recognizes repeat offenders (GDPR-compliant), and alerts staff instantly.
Need for advanced AI expertise
The client required specialized computer vision engineers to build custom models that could accurately detect specific theft behaviors in busy stores.
Hardware scalability
They struggled to make the software run smoothly on affordable edge devices while processing video from 20+ cameras simultaneously.
Real-time performance
The client needed a complex architecture that could deliver alerts in milliseconds without crashing the system.
Strict privacy compliance
Operating in the UK, they needed a partner who could ensure facial recognition and data storage met strict GDPR standards.
Solutions we implemented
Geniusee engineered a comprehensive, custom AI platform designed to transform legacy CCTV infrastructure into an active security asset. We delivered an end-to-end solution that included core behavior-detection models and a user-friendly admin interface. Our team customized the solution to fit the client’s specific operational workflows.
1. Development of core system capabilities
- Custom AI & edge integration. We developed proprietary behavior-detection components that run directly on edge devices, enabling the system to work seamlessly with the client’s existing in-store camera infrastructure.
- Real-time stream processing. The system processes camera streams in real-time, identifying shoplifting attempts and recognizing banned individuals (via facial recognition) with immediate alerting capabilities.
- Centralized monitoring dashboard. We built a full-featured internal web application that allows security teams to review events, confirm detections, and manage the system.
- Advanced data handling. The solution includes automated watermark generation for evidence integrity and integration with vector storage to manage facial data for future detection enhancements.
2. Customized approach
To meet the client’s performance expectations and long-term goals, we tailored our technical approach:
- Multi-stage processing pipeline. To ensure instant visibility, we designed a split pipeline. Essential event data is sent via Telegram alerts in near real-time, while computationally intensive tasks (such as post-processing and storage) are performed in the background.
- Dynamic scalability. We implemented a dynamically scalable pipeline on the edge computer. Although initially developed and tested on a single camera, the architecture was designed to process events from multiple cameras in parallel after release.
- Frontend-contract-first development. We utilized domain-driven design with full mock data on the frontend. This allowed us to build and refine the user interface functionality in parallel with backend development, significantly accelerating the timeline.
Flexible deployment phases: We designed the system to support a smooth transition between different project phases (e.g., from MVP to exhibition), ensuring the architecture remained stable despite evolving priorities.
3. Quality assurance & testing
We went beyond standard software testing to ensure physical reliability:
- In-person physical testing. We validated action and face detection models by setting up physical edge devices and on-site cameras to mimic real-world conditions.
- Comprehensive test suite. Our QA team prepared and executed a full test suite covering backend, web front-end, and AI model accuracy.
- Integration testing. We performed functional and change-related testing at both system and integration levels to guarantee end-to-end stability.
Telegram notifications
We integrated fast, reliable Telegram notifications that trigger moments after new events occur. This feature includes interactive actions that allow operators to classify events, confirm or reject detections, and respond quickly directly from the chat without opening the admin panel.
Real-time dashboard
The system features a centralized dashboard for monitoring event flow and reviewing media in real time. It provides operators with complete visibility, allowing them to check processing status and view the full alert history at a glance.
Shoplifting & face detection
The platform performs real-time shoplifting detection alongside face detection to recognize banned individuals. We optimized this for speed by implementing immediate alerts with parallel media saving, ensuring that the notification process is never delayed by data storage.
Scalable architecture
All components are designed for horizontal scaling as event volume grows, with a clear separation between ingestion, processing, alerting, and storage services. The system supports multiple enrichment steps (metadata extraction, tagging, filtering) and is easily extendable to integrate future business logic or AI models.
Enhanced visibility & control
We delivered a custom admin panel that centralizes system monitoring, event review, and user management. This interface reduces manual workload and empowers the client with faster, data-driven decision-making capabilities.
Rapid response via interactive alerts
Instant Telegram notifications provide immediate situational awareness. We added interactive controls that allow staff to classify events, confirm detections, or reject false alarms directly within the chat app.
Optimized event processing
Multi-stage pipeline ensures that critical data surfaces immediately while heavy processing completes in the background. This approach improved system responsiveness without sacrificing data completeness or evidence integrity.
Future-ready scalability
The architecture was designed for horizontal growth. As event volumes increase, the system scales seamlessly without the need for re-architecting, ensuring long-term performance stability.
Foundation for advanced AI
By integrating vector storage and structured event data collection, we established a reliable foundation for the future training of advanced AI models, including facial recognition clustering and predictive event ranking.
Accelerated planning
The introduction of a dedicated researcher role and the use of mock data allowed for faster planning and refining, enabling successful demos even when backend data was not yet fully available.
Transitioning from research to working AI
Challenge: The initial AI model and incoming data were inconsistent and unsuited for heavy, real-world workloads.
Resolution: We replaced the research-phase model with an optimized engine and built a “data safety net” to automatically correct errors and fill in missing information, ensuring the system remains stable.
Scaling hardware for 20+ camera streams scalability
Challenge: Processing video from over 20 cameras simultaneously required significant power and verified stability.
Resolution: After deep technical analysis, we sourced specific edge devices capable of multi-stream processing and validated the full flow by scaling from 2 to 20+ physical cameras.
Balancing speed with heavy processing
Challenge: Real-time alerts were often delayed by computationally heavy background analysis.
Resolution: We split the workload: lightweight tasks trigger instant alerts, while intensive analysis happens in the background, ensuring staff receive notifications in milliseconds.
Ensuring reliable, real-time communication
Challenge: Initial Telegram alerts were unreliable due to internet glitches and inconsistent data.
Resolution: We integrated automatic “retry” logic and delivery tracking to ensure alerts reach security staff every time, even during network fluctuations.
Adapting to specific store environments
Challenge: AI performance often dropped when faced with unique store lighting, angles, or new exhibition layouts.
Resolution: We replicated exact store conditions with a dedicated office camera and used on-site data to quickly retrain models, ensuring the AI understood every specific location.
Navigating compliance & remote deployment
Challenge: Late-stage GDPR changes and hardware access issues threatened the launch timeline.
Resolution: We prioritized privacy-first development and used detailed system logs to fix code locally, ensuring a legally safe and successful remote deployment.
