SentinelAI
Hybrid Edge–Cloud Violence Detection Architecture
A production-oriented AI surveillance system that combines lightweight MobileNetV2 edge inference with Vision-Language Model verification to intelligently detect violent events while minimising cloud computation.
Overview
SentinelAI is my final-year Computer Science dissertation project at Coventry University. It explores a hybrid Edge–Cloud AI architecture for intelligent CCTV violence detection.
Instead of sending every frame to the cloud, the system first performs lightweight edge inference using MobileNetV2. Only suspicious events are escalated to a Vision-Language Model for multimodal verification — reducing computational cost while maintaining reliable detection.
The Problem
Traditional cloud-based surveillance systems process every frame remotely, failing across four critical dimensions.
High Computational Cost
Every frame processed remotely at full cloud inference cost, regardless of content.
Increased Latency
Constant round-trip cloud calls on every frame introduce unacceptable detection delays.
Poor Scalability
Uniform processing fails to scale efficiently across dense camera grids.
Unnecessary Inference
95%+ of frames contain no violence and should never consume cloud resources.
Solution
A two-stage intelligent pipeline: lightweight edge inference handles the volume, cloud VLM handles the complexity — only when required.
Camera Feed
Continuous CCTV frame capture
Edge AI — MobileNetV2
Lightweight on-device violence classification
Suspicion Threshold
Confidence gate — classify or escalate
Only Suspicious Frames
Selective escalation to cloud — not every frame
VLM Verification
OpenAI Vision API — multimodal semantic analysis
Final Decision
Combined edge + VLM classification output
Logging & Analysis
Event dashboard and structured audit trail
Key Features
Hybrid Edge–Cloud AI Architecture
Two-stage intelligent pipeline balancing compute efficiency with detection accuracy across the full inference chain.
MobileNetV2 Edge Inference
Depth-wise separable CNN deployed on-device for high-throughput, low-latency first-pass violence detection.
Vision-Language Model Verification
OpenAI Vision API provides multimodal semantic understanding for contextual verification of suspicious events.
Selective Escalation Pipeline
Only frames exceeding the suspicion threshold reach the cloud — dramatically reducing inference cost.
Deployment-Oriented Evaluation
System assessed against balanced real-world datasets with production performance constraints in mind.
Production Web Demonstration
Full Next.js interface deployed on Vercel — live and publicly accessible for end-to-end demonstration.
Responsive Full-Stack Interface
Production-quality UI built with TypeScript and Tailwind, optimised for desktop and mobile viewports.
Technologies
Project Outcome
Final-Year Dissertation — Coventry University
A complete, production-deployed AI system demonstrating that hybrid edge-cloud architectures can significantly reduce computational cost while maintaining detection reliability.
Live Experience
Experience SentinelAI
Explore the live production deployment or watch the full project demonstration to see the hybrid architecture in action.