Inspiration
This project was inspired by the need to filter, verify, and transform live disaster data into reliable, structured intelligence that authorities and responders can act upon immediately.
What it does
Combine Machine Learning filtering with Gemini AI reasoning to convert noisy real-time data into verified disaster reports.
How we built it
Architecture Overview
The system follows a hybrid pipeline:
Live Data Sources → ML Filter → Verification Pipeline → Gemini API → Structured Report Generator → Frontend Dashboard
Backend (Flask API Server)
Built using Python + Flask
Created REST endpoints:
/api/analyze
/api/disaster-summary
Stored API keys securely in a .env file
Used python-dotenv to load environment variables
Integrated Gemini API for structured disaster analysis
The backend:
Filters raw disaster input
Verifies structured parameters (location, severity, affected count)
Sends refined prompt to Gemini
Returns AI-generated structured report
ML Filtering Layer
Since traditional ML models were not pre-trained:
Implemented rule-based filtering (keyword + threshold detection)
Structured validation (e.g., displacement count > 1000 triggers severity flag)
Confidence scoring logic
This acts as a pre-LLM validation firewall before calling Gemini.
Gemini API Integration
Gemini is used for:
Situation assessment
Risk classification
Priority level generation
Resource recommendation
Structured executive summary
Gemini does reasoning, not filtering. The ML layer ensures only validated data reaches the model.
Challenges we ran into
Handling Real-Time Data Noise
Raw disaster information is inconsistent and unstructured. Solution: Built a validation and filtering layer before AI processing.
Controlling AI Output
LLMs can generate unpredictable responses. Solution: Used structured prompts requesting JSON-formatted output for consistency.
What we learned
The final system:
Accepts live disaster data
Filters and verifies it
Uses Gemini AI for intelligent reasoning
Generates structured, actionable reports
Displays results instantly on a dashboard
It demonstrates a hybrid ML + LLM architecture suitable for real-world emergency intelligence systems.
Log in or sign up for Devpost to join the conversation.