Inspiration
The core inspiration for Sentinel AI comes from a simple, devastating truth: "The disaster isn't the problem. The blind spot before it is."
India lost $79.5B (approx. ₹6.6 Lakh Crore) to climate disasters between 1998 and 2017 (UNDRR). With over 100 crore Indians living in disaster-prone zones, we realized that response time measured in hours must become seconds. We noticed that every hazard—whether flood, cyclone, or fire—shares the same points of failure: fragmented departments coordinating manually, field reports taking hours to verify, language barriers, and total connectivity collapse when mobile networks fail. We built Sentinel AI to eliminate this blind spot and transform disaster management from a reactive scramble into a proactive, coordinated response.
What it does
Sentinel AI is a Predictive Risk Intelligence & Coordinated Response System that covers the full disaster lifecycle:
BEFORE (Digital Twin + Prediction): Live data from satellites, rivers, weather, and government sources feed a real-time city risk model. AI flags high-risk zones on an H3 risk grid and fires alerts before the disaster peaks. It converts standard forecasts into precise hazard windows (start, peak, ease) per grid cell.
DURING (Detect, Verify, Dispatch): Citizens can report hazards via a PWA, a zero-internet AI Calling Agent, or a one-tap SOS. A 4-signal AI scoring system (Image 25%, Heatmap 25%, User 25%, Climate 25%) instantly verifies reports, blocks screen-recapture fakes, and dispatches nearby volunteers via WhatsApp in under 3 seconds.
AFTER (Rebuild, Reconnect, Recover): Our LifeLine P2P marketplace matches survivors' needs directly with donors' supplies, closing the gap that institutional aid can't reach fast enough.
Core Accessibility: The system is powered by OpenAI with a Nemotron fallback, supports 6 languages (English, Hindi, Telugu, Tamil, Malayalam, Kannada), and functions entirely offline using SMS/Voice fallbacks and PWA sync.
How we built it
We built a robust, highly scalable architecture designed to survive real-world disasters:
Backend & Database: Python, Flask, Docker, PostgreSQL + PostGIS (for spatial indexing), SQLite, and Uber H3 for hexagonal risk grids.
AI & LLM: OpenAI GPT and NVIDIA NeMoTron (as a reliable fallback) for verification. We use LangGraph to orchestrate the Prediction Agent's hazard windows.
Frontend: A Progressive Web App (PWA) built with HTML5, CSS3, JavaScript, and Bootstrap 5. It features Service Workers for offline-first reporting and automatic syncing.
Integrations & APIs: Twilio (for WhatsApp dispatch and AI Voice Calls), Leaflet.js/MapLibre GL, Open-Meteo, NASA GIBS, RainViewer, CWC, and Mapillary.
Jobs & Messaging: APScheduler and REST API layers for background task management.
Challenges we ran into
Connectivity Collapse: Mobile networks often fail during disasters. We solved this by building an AI Calling Agent that works on any basic phone (no data needed) and an offline PWA that syncs when the internet returns.
AI Verification Accuracy: False positives can trigger incorrect responses. We implemented a multi-source verification system using location, weather, report history, and image analysis.
Evidence Integrity: People panic and submit duplicates. We built a system to detect screen-recaptures and duplicate reports before they reach a human reviewer, ensuring no second team is dispatched to the same spot.
Responder Safety & Privacy: Volunteers entering hazardous zones are protected via geo-fenced dispatch and hazard-aware routing. To handle citizen PII (photos/voice), we are planning DPDP Act 2023-compliant self-hosted processing and India-region data residency as a pre-scale milestone.
Accomplishments that we're proud of Achieved a verified dispatch time of ≤ 3 seconds from verification to volunteer notification.
Built a working Digital Twin prototype that rescans city zones every 5 minutes.
Created a 4-signal AI scoring system that successfully blocks duplicate and fraudulent reports while pushing critical alerts (Score ≥85%) to the top of the queue.
Developed a fully functional offline-first reporting system with voice call and SMS fallbacks.
What we learned
We learned that technology in disaster management must be resilient, not just smart. Building fallback systems—like Nemotron for AI and SMS/Voice for data—is just as important as the core prediction algorithms. We also learned how to utilize open data sources (NASA, IMD, CWC) to create a unified, real-time command view for municipal commissioners, proving that bridging fragmented departments is possible through a shared technological grid.
What's next for Sentinel AI
We have a clear roadmap to scale from a working prototype to a district-wide deployment:
Next Steps: Implementing PostGIS spatial indexing for fast radius queries, WebSocket push alerts to replace notification polling, scaled APScheduler jobs for WhatsApp/SMS fan-out, and scheduled hazard-window recomputation by the Prediction Agent.
Later Phases: Integrating live IoT river-gauge and drainage sensors, launching native mobile apps with background geofencing, and utilizing satellite SAR inundation mapping for post-event damage assessment.
Pilot & Scale: We plan a 12-month pilot starting with Hyderabad urban flooding (GHMC), moving to Bengaluru, then a coastal district (cyclone/surge) before district-wide scale-up. We aim to integrate with NDMA/SDMA SACHET incident-exchange standards and the Smart Cities Mission.
Built With
- apscheduler
- css3
- docker
- flask
- html5
- javascript
- langgraph
- leaflet.js
- maplibre-gl
- nasa-gibs
- nvidia-nemotron
- open-meteo-api
- openai
- postgis
- postgresql
- pwa
- python
- render
- service-workers
- sqlite
- twilio
- uber-h3
