AI-Powered Smart Surveillance System
A real-time surveillance system that automates car and human identification using Vision-Language and Large Language Models, cutting manual monitoring effort by 90%.
Challenge
Manual video surveillance doesn't scale: identifying vehicles and people across live camera feeds takes constant human attention, and the workload grows with every added stream. The client needed a system that could watch at scale without a proportional increase in headcount.
Approach
We led a team of three to design and build a surveillance system combining Vision-Language Models with Large Language Models (GPT-4, Llama 3.1) for identification and reasoning over live footage. RTSP streaming brought camera feeds into the pipeline, Apache Kafka handled real-time event distribution, and CUDA-accelerated inference kept latency low enough for live monitoring, all deployed on Kubernetes for scale. The architecture, client pitching, and stakeholder management ran alongside the engineering.
Outcome
The system cut manual identification effort by 90%, replacing constant human monitoring with automated detection and alerting. The result is a scalable, AI-driven surveillance platform built for production camera loads rather than a demo.