SmartCCTV — AI Surveillance & Monitoring Platform
Turn any existing CCTV camera into an intelligent monitoring system — no new hardware, just plug in and the camera starts thinking.
- Deployed across 20+ cameras at pilot sites
- 10,000+ daily face-recognition events
- Live React monitoring dashboard
What needed solving
Most CCTV systems in India just record — they don't understand what they're seeing. Upgrading to 'smart' surveillance usually means ripping out existing cameras and buying expensive proprietary hardware, which is a non-starter for most schools, offices, and small businesses.
What I built
I built a software-only AI layer that sits on top of any existing IP camera setup and adds real-time intelligence — face recognition, attendance tracking, and unknown-person alerts — without replacing a single piece of hardware. Connect a camera to the system, and it instantly becomes intelligent.
Architecture
Video ingestion
Pulls live RTSP/IP streams from existing cameras, no special hardware required.
Detection
YOLOv9 runs real-time object/person detection on each video frame.
Recognition
ArcFace generates facial embeddings and matches them against a known-faces database for identity recognition.
Event pipeline
Recognized/unrecognized faces trigger structured events (attendance log, alert, etc.), processed through a FastAPI backend.
Storage & alerts
PostgreSQL stores attendance and event logs; Redis handles real-time alert queuing and caching for fast lookups.
Live dashboard
A React-based monitoring dashboard shows live camera feeds, face-recognition events, and alerts in real time.
Key challenges solved
Real-time performance at scale
Processing 10,000+ daily face-recognition events across 20+ cameras meant optimizing the detection-to-recognition pipeline so it could run continuously without lag, balancing accuracy against compute cost.
Zero hardware dependency
Designing the system to work with whatever IP cameras a site already has — rather than requiring specific proprietary hardware — meant building a flexible ingestion layer that handles varying stream qualities and protocols.
False positive control
Unknown-person alerts need to be reliable — too many false alarms make the system useless. Tuning ArcFace's similarity thresholds and adding consistency checks across multiple frames reduced false positives significantly.
Live dashboard responsiveness
Showing real-time events from 20+ camera feeds simultaneously without overwhelming the frontend or backend required efficient event batching and Redis-backed caching.
What it shipped
Other projects
Have a system that needs to ship?
I build production-grade AI — voice, vision, and full-stack. Open to senior AI engineering and founder roles.