Inspiration
The disaster isn't the problem. The blind spot before it is.
India lost $79.5 billion to climate disasters between 1998 and 2017 (UNDRR), and with over 100 crore Indians living in disaster-prone zones, response times measured in hours must become seconds. We were inspired to build Sentinel AI because every hazard—flood, cyclone, surge, or fire—shares one fatal failure: fragmented departments, slow verification, language barriers, and connectivity collapse mean nobody sees it coming. We wanted to build a system that not only predicts the risk but coordinates the response seamlessly.
What it does
Sentinel AI is a Predictive Risk Intelligence & Coordinated Response System that covers the full disaster lifecycle on a unified H3 risk grid.
Before (Prediction): We use a Digital Twin to ingest live data from satellites (ISRO/NASA), weather APIs, and river gauges. Our AI flags high-risk zones and fires geo-fenced alerts, predicting the exact hazard window (start, peak, ease).
During (Detect, Verify, Dispatch): Citizens can report hazards via an Offline PWA, a 1-tap SOS, or our AI Calling Agent (which works without internet in 6 languages). Our 4-signal AI verification engine (Image 25%, Heatmap 25%, User 25%, Climate 25%) checks for duplicates and screen-recapture fakes. If a report scores ≥85%, it triggers a priority dispatch to nearby volunteers via WhatsApp in under 3 seconds.
After (Recover): Our "LifeLine P2P" matches survivors' needs with donors' supplies automatically, closing the gap institutional aid can't reach fast enough. Every resolution requires GPS + photo proof before closure.
How we built it
We built the platform as a Progressive Web App (PWA) to ensure offline-first functionality.
Backend: Python, Flask, PostgreSQL + PostGIS for spatial indexing.
AI & NLP: We use OpenAI GPT for primary verification and hazard forecasting, with NVIDIA Nemotron as a fallback so scoring never fails. LangGraph orchestrates the prediction agents.
Voice & Dispatch: Twilio API for WhatsApp dispatch and the AI Call Agent, supporting English, Hindi, Telugu, Tamil, Malayalam, and Kannada.
Geospatial: Leaflet.js, MapLibre GL, and Uber's H3 hexagonal grid to power the Digital Twin's risk scoring.
Challenges we ran into
Connectivity Collapse: Mobile networks usually fail during disasters. We had to architect an offline-first PWA that syncs when connection returns, alongside voice and SMS fallback mechanisms.
Data Reliability & Fraud: We needed to stop fake reports (like screen-recaptured photos) from triggering unnecessary deployments. We solved this by implementing a rigorous 4-signal AI scoring system that holds anything under a 50% confidence score for human corroboration.
Language Barriers: Ensuring alerts reach non-English communities was critical, requiring robust NLP translation and text-to-speech pipelines for 6 regional Indian languages.
Accomplishments that we're proud of ≤3 Sec Dispatch: We achieved verification-to-volunteer notification in under 3 seconds.
Zero-Internet Reporting: Our AI Calling Agent successfully logs and verifies reports via voice call, ensuring no one is left behind.
Evidence Integrity: Successfully implemented algorithms to detect screen-recapture photos, blocking fraudulent reports before they reach a supervisor.
Live Prototype: Our Digital Twin is currently live for 2 cities, and we are monitoring 26+ regions with our forecast models.
What we learned
We learned the immense value of building resilient systems. We discovered that the most advanced AI is useless if the user cannot connect to the internet. Designing fallback models (OpenAI to Nemotron) and alternative input methods (Voice AI vs. App) taught us how to prioritize accessibility and reliability in high-stress environments.
What's next for Sentinel AI
Next: PostGIS spatial indexing for fast radius queries at city scale, WebSocket push alerts, and scaled APScheduler jobs for SMS fan-out.
Later: Ingesting live IoT river-gauge data, Native mobile apps with background geofencing, Satellite SAR inundation mapping for post-event damage assessment, and full DPDP Act 2023 compliance with India-region data residency.
Pilot: A 12-month rollout starting with Hyderabad urban flooding (GHMC), scaling to Bengaluru, and then coastal districts.
Log in or sign up for Devpost to join the conversation.