DisasterLens — AI Emergency Intelligence
AI emergency coordination with Gemma 4: triage, de-duplicate, and map Nepali & English disaster reports on one responder dashboard.
DisasterLens is an AI-powered emergency coordination platform that uses Gemma 4 to analyze disaster reports, extract critical information, prioritize urgent incidents, and visualize affected locations on an interactive map — all through a unified dashboard designed to support faster, more informed disaster response.
THE PROBLEM: In fast-moving emergencies, responders drown in unstructured reports — social media posts, phone calls, SMS — with no consistent way to extract location, severity, or trapped-person signals. Duplicate reports fragment situational awareness and delay triage.
WHAT WORKS:
• Structured extraction: incident type, severity (1–5), location text, people trapped, road blocked, hazards, vulnerable groups, needs — returned as typed JSON
• Nepali / English support: Devanagari, Romanized Nepali, and English reports analyzed with language detection
• Duplicate grouping: heuristic Jaccard similarity on text + location, with optional LLM judge; incidents share one lifecycle
• Urgency scoring with safety net: 0–100 triage score from severity, casualties, vulnerable groups, trapped people, road blockages, corroborating reports; clause-aware keyword overrides for “trapped” in Nepali and English
• Lifecycle tracking: new → under_review → verified → response_in_progress → resolved, with actor-attributed history; status updates apply to entire duplicate groups
• Situation briefing: LLM-generated summary from up to 30 distinct incidents, evidence-only, fails safely (HTTP 503) if the model is unavailable
Runs locally with a stub mode (DISASTERLENS_STUB=1) for development without API costs; the open-weight model endpoint (NVIDIA-hosted Gemma 4) keeps inference accessible. Backend is pure Python/FastAPI/SQLite with a Leaflet map dashboard; deployable via the included Render blueprint.
Built with
AI tools
Prize Categories Entered
- Submitted
Best Use of Gemma 4
Gemma 4 drives DisasterLens's entire analysis pipeline: it classifies and extracts structured incident data (type, severity 1-5, location, people trapped, hazards, needs) from raw disaster reports, then generates the situation briefing for responders. One multilingual prompt handles Devanagari, Romanized Nepali and English without separate models, and the open-weight model is served through an OpenAI-compatible endpoint (NVIDIA NIM by default, with a key-free stub mode for development).
- Event prizeSubmitted
Best use of open source AI for Real-World Impact
Nepal knows the problem: responders drowning in unstructured reports during floods and earthquakes. DisasterLens turns them - SMS, calls, social posts, in Nepali or English - into a de-duplicated, prioritized incident map with 0-100 triage scoring, clause-aware trapped-person keyword overrides, and a lifecycle from report to resolution. AI is the core: remove Gemma 4 and nothing works. Anyone can run it today: the repo is public and MIT-licensed, installs with pip, and ships a stub mode that needs no API key.
