The Challenge

A major U.S. retail chain was experiencing unauthorized after-hours access in multiple store locations. Their existing camera setup lacked context, had no real-time alerts, and required manual arm/disarm procedures—leading to frequent false positives during business hours and missed events overnight.

They needed an automated, real-time system that could:

• Detect intrusions precisely when stores were closed
• Capture visual context using smart cameras
• Alert staff instantly—without triggering false alarms during store hours

Our Approach

We built a hybrid edge-cloud intrusion detection system using door sensors, motion-activated cameras, and AI-based face detection—coordinated by AWS services and integrated with Slack for real-time alerts. The system could:

• Arm/disarm automatically based on store hours
• Capture images 5 seconds before and after any suspicious motion
• Use local AI for face detection (with no cloud data leaks)
• Deliver alerts in <5 seconds with annotated images

What We Delivered

Intelligent Edge Device

• Raspberry Pi with camera, motion sensor, and door status detection
• Local image buffer built with OpenCV
• DeepStack AI for on-device face detection
• Synced timestamps to align sensors and cameras perfectly

Serverless & Scalable Backend

• AWS IoT Core for secure MQTT-based communication
• AWS Lambda for image event processing and Slack notifications
• AWS DynamoDB for storing logs and metadata
• Spring Boot APIs for admin controls
• React Dashboard with role-based access via AWS Cognito

Automation & Security

• System auto-arms/disarms via AWS Scheduler
• JWT-secured APIs using Spring Security + Cognito
• IAM and X.509-based device security

‍

Tech Stack Highlights

‍

Results After Deployment

‍

Managers also began using the admin dashboard daily to adjust store schedules and review logs, increasing engagement and control.

‍

Lessons We Learned

• AWS IoT Core greatly simplified device communication at scale
• Local AI processing avoided privacy risks and cloud latency
• OpenCV buffering was key to capturing images just before/after each event
• Slack was perfect for MVP alerts, though advanced notification logic is needed for production
• Avoiding observability in MVP helped speed up delivery—but made debugging trickier

‍

What’s Next

We're working on scaling this to 100+ store locations with new features:

‍

Want to build something smart and secure like this?

👉 Let’s talk.

‍