AI CCTV Control Room Operator

dashbaord

Inspiration

We wanted to build a real-time AI system that helps security operators detect and understand incidents instantly, with explainable alerts and instant video evidence — all without complex setup or heavy infrastructure.

What It Does

Detects incidents in real time from live camera feeds using Google Gemini.
Slices the stream into short MP4 clips and attaches the relevant snippet to each detected event.
Streams events to the dashboard via Server-Sent Events (SSE) with concise, explainable summaries.
Allows operators to preview clips instantly and review recent alerts.
Provides backend APIs to start/stop monitoring and view system status.

How We Built It

Backend: FastAPI service with lifecycle management, CORS, and a video pipeline that chunks streams into video_chunks/. Event detection is handled by Gemini, with optional persistence to PostgreSQL when DATABASE_URL is configured.
Realtime: SSE endpoint streams new events to clients without polling.
Frontend: Next.js (App Router) app that renders the live monitoring UI, subscribes to SSE, and plays event-specific clips.
Data Model: Each event includes timestamp, event code, description, explainability text, and clip URL.
Operations: Simple endpoints — /start, /stop, /status, /events, /events/id/{event_id}, and /video?filepath=....
Latency Optimization: We target sub-5s perceived delay by overlapping video chunking and analysis. If chunk size = L and analysis time = A, total delay ≈ max(L, A) + δ (network/IO). We tune L to balance quality, latency, and cost.

Challenges

Reducing latency while maintaining consistent video quality and timely event detection.

Accomplishments

Achieved a clean end-to-end flow from camera to explainable alert with one-click clip playback.
Implemented smooth, real-time updates using SSE with client-side deduplication.
Designed a graceful service lifecycle with safe startup/shutdown and optional database persistence.

What We Learned

In operator workflows, clarity beats volume — short, high-signal summaries with precise clips outperform verbose analytics.
SSE is ideal for one-way real-time alerts — simpler and more efficient than WebSockets for this scenario.
Clear architectural seams (typed schemas, defined endpoints) make AI components easy to swap without breaking the UI.

What’s Next

Multi-Camera Orchestration: Support for multiple feeds with per-camera policies and prioritization.
Edge Acceleration: On-premise inference to reduce latency and cost.