CERTIS — Predictive Agentic Security Advisor

Our submission for NAISC 2026. It's a security operations dashboard that uses Claude to analyse live camera feeds, audio, and access control logs to detect threats and surface recommended actions to on-site officers.

Three agents (vision, audio, log) run in parallel on incoming data. A coordinator agent synthesises their findings and raises an alert if confidence clears the threshold. Alerts hit the dashboard in real time over WebSocket.

Tech Stack

• Backend: FastAPI + LangGraph + SQLite
• AI: Anthropic Claude (vision + text)
• Frontend: React + TypeScript + Tailwind CSS + Vite

Getting Started

You'll need Python 3.10+, Node.js 18+, and an Anthropic API key.

Backend

cd backend

python -m venv venv
source venv/Scripts/activate      # Windows (bash)
# source venv/bin/activate         # macOS / Linux

pip install -r requirements.txt
uvicorn main:app --reload

Runs at http://localhost:8000. Swagger docs at /docs.

Frontend

Open a second terminal, then run:

cd frontend
npm install
npm run dev

Runs at http://localhost:5173.

Environment Variables

Create a backend/.env before starting the backend:

ANTHROPIC_API_KEY=sk-ant-...
CLAUDE_MODEL=claude-sonnet-4-6
DATABASE_URL=sqlite+aiosqlite:///./security.db
FRAME_ANALYSIS_INTERVAL=3
ALERT_COOLDOWN=30

Variable	What it does	Default
ANTHROPIC_API_KEY	Your Anthropic API key	required
CLAUDE_MODEL	Which Claude model to use	claude-sonnet-4-6
DATABASE_URL	Database connection string	SQLite local file
FRAME_ANALYSIS_INTERVAL	Only analyse every Nth camera frame	3
ALERT_COOLDOWN	Seconds before the same alert type can fire again	30

How It Works

Camera (WebSocket) ──► Vision Agent ──┐
Audio  (REST)      ──► Audio Agent  ──┼──► Coordinator ──► Alert ──► WebSocket broadcast
Logs   (REST)      ──► Log Agent   ──┘

The Vision Agent checks CCTV frames for physical threats. The Audio Agent handles intercom/microphone transcripts. The Log Agent applies rules to access control and alarm events — no LLM, just rules. The Coordinator takes all three and decides whether to raise an alert and what officers should do next.

Alerts only fire if confidence is ≥ 0.65. If the Claude API goes down, the coordinator falls back to rule-based logic.

API Endpoints

Method	Endpoint	Description
GET	/api/health	Health check
GET	/api/alerts	Get stored alerts
POST	/api/audio	Send a transcript for analysis
POST	/api/logs/event	Send an access control / alarm event
POST	/api/feedback	Officer confirms or dismisses an alert
WebSocket	/api/ws/alerts	Real-time alert stream
WebSocket	/ws/camera	Camera frame ingestion

Project Structure

├── backend/
│   ├── main.py
│   ├── agents/
│   │   ├── vision_agent.py
│   │   ├── audio_agent.py
│   │   ├── log_agent.py
│   │   └── coordinator.py
│   ├── core/
│   │   ├── config.py
│   │   └── models.py
│   ├── routers/
│   └── graph.py
└── frontend/
    └── src/
        ├── components/
        └── hooks/

Notes

• Microphone transcription only works on Chrome or Edge (Web Speech API limitation)
• The demo panel lets you simulate events (fire alarm, forced entry, panic button, etc.) without real hardware
• Webcam access is required for the camera feed