Sentinel AI
Multi-agent AI that predicts hazard windows, verifies citizen reports in seconds and dispatches volunteers, even with no internet.
Built at AgentX Hackathon 2026
Team TECH MAX: varun, Manasvi Gudey
Built with Python, Flask, PostgreSQL, PostGIS, SQLite, LangGraph, OpenAI GPT, NVIDIA Nemotron, Qwen3, Twilio, APScheduler, Uber H3, Leaflet.js, MapLibre GL, Open-Meteo, RainViewer, Chart.js, Bootstrap 5, JavaScript, PWA, Docker, Render
About the project
PROBLEM
Disaster response in India fails in the blind spot before and during the event: departments coordinate manually with no shared picture, field reports take hours to verify and get duplicated, alerts miss non-English speakers, and mobile networks collapse when they are needed most. India lost $79.5B to climate disasters between 1998 and 2017 (UNDRR).
SOLUTION
Sentinel AI is a multi-agent disaster response coordinator covering the full lifecycle on one shared H3 risk grid. It is flood-first and extensible to cyclone, surge and fire.
- Before: a real-time city Digital Twin scores hazard x vulnerability, and a Prediction Agent turns forecasts into per-cell hazard windows (start, peak, ease).
- During: citizens report through the PWA, a one-tap voice SOS, or an AI Calling Agent that works on any phone with no mobile data. Reports are verified, deduplicated and dispatched automatically.
- After: LifeLine P2P matches survivors' needs with donors' supplies.
HOW IT WORKS
1. Risk Monitoring Agent rescans every zone every 5 minutes across 20+ weather, satellite, river and government data sources, and raises alerts before the hazard peaks.
2. Verification Agent scores each report on 4 signals at 25% each (image analysis, heatmap match, climate alignment, user reliability). It flags screen-recaptured photos and matches duplicates within ~5.5 km / 24 h.
3. Prioritisation: a score of 85% or above goes to priority dispatch, 50-84% to official review, and below 50% is held for corroboration.
4. Dispatch Agent sends geo-fenced alerts and notifies the nearest volunteer on WhatsApp/SMS with hazard-aware routing. A case closes only with GPS + photo proof.
5. OpenAI leads verification with NVIDIA Nemotron as fallback; LangGraph orchestrates the agents. Supports English, Hindi, Telugu, Tamil, Malayalam and Kannada.
Five role-based portals (citizen, volunteer, supervisor, analyst, commissioner) share one timestamped audit trail.
IMPACT
- 3 seconds or less from verification to volunteer notification (tested)
- 26+ regions under forecast monitoring; Digital Twin live in 2 cities (Hyderabad, Bengaluru)
- 6 languages and 3 reporting channels, including zero-internet voice calls
- No duplicate deployments: matched reports are held instead of sending a second team
- Geo-fenced alerts warn only the people actually at risk
- Pilot cost of roughly Rs 8,000-10,000 per city per month, with no dedicated hardware