Multimodal AI System for Real-Time Disaster Analytics & Response.
Combines NLP, Computer Vision, Agentic RAG, and Predictive Modeling to transform chaotic data into actionable intelligence.
During a disaster, decision-makers are overwhelmed by unstructured data. DisasterInsight AI is an end-to-end platform that unifies diverse AI disciplines to solve this:
- Sees damage using Computer Vision.
- Reads crisis reports using NLP.
- Predicts future risks using Time-Series Forecasting.
- Reasons and plans using a Generative AI Agent with access to official protocols (RAG).
Click the thumbnail below to watch a short demo of the platform.
Dashboard:
Auto-Triage of Disaster Imagery:
Forecasts:
Smart Chat Assistant (RAG based):
This is not just a dashboard; it is an Orchestration of 5 AI Modules:
- Brain: Powered by Google Gemini 2.0.
- Tools: The agent can autonomously call the Risk Model, Forecast Model, or search the database based on user queries.
- RAG (Retrieval Augmented Generation): Queries a vector database (ChromaDB) of official PDF safety protocols to provide verified advice, eliminating hallucinations.
- Model: Fine-tuned MobileNetV2 (served via ONNX Runtime for low latency).
- Function: Classifies uploaded images (e.g., "Major Damage", "Flood", "Fire") and automatically assigns a Triage Priority (Critical/High/Low).
- Model: DistilBERT Transformer.
- Function: Classifies social media streams into 10 humanitarian categories (e.g., "Rescue Needed", "Infrastructure Damage") in real-time.
- Global Forecast: Uses Prophet to predict long-term seismic trends.
- Regional Impact: Uses XGBoost to predict the probability of high-fatality events in specific high-risk zones.
+------------------------+ +---------------------------+ +--------------------------+
| | | | | |
| React Frontend | ---> | FastAPI Backend | ---> | 4x AI / ML Models |
| (Tailwind, Chart.js, | | (Python, Uvicorn) | | (Transformers, XGBoost, |
| Mapbox GL) | | | | Prophet, ONNX CV) |
| | | | | |
+------------------------+ +---------------------------+ +--------------------------+
The system follows a decoupled, microservices-ready architecture, integrating AI models and agentic workflows.
- Frontend: React, Tailwind CSS, Chart.js, Mapbox GL
- Backend API: Python, FastAPI, Uvicorn, Docker
- AI & ML Models: PyTorch, Transformers(DistilBERT), Scikit-learn, XGBoost, Prophet, ONNX CV Models
- Data & Storage: Pandas, Jupyter, ChromaDB (for RAG)
- Agentic Workflow: Gemini AI Agent orchestrating model calls & retrieval
- Agentic Workflow: Gemini AI Agent orchestrating model calls & retrieval
This repository includes comprehensive documentation for both end-users and developers.
-
Project Summary Report: A high-level overview of the project's objectives, methodology, and key results.
-
Model Performance Reports: In-depth analysis of each of the four AI models, including metrics, confusion matrices, and feature importance.
-
Exploratory Data Analysis (EDA): Reports on the initial data analysis that informed our modeling strategies.
-
Dashboard User Guide: A detailed walkthrough of all features in the live web application.
This monorepo contains all the code and assets for the DisasterInsight AI platform. Here's a guide to the key directories:
| Folder | Description |
|---|---|
| data | Instructions and links to download the datasets used for model training. (Data files are not included). |
| disaster-insight-api | The high-performance FastAPI backend that serves the AI models. |
| disaster-insight-frontend | The production-grade React frontend application. This is the main user interface. |
| legacy_streamlit_ui | The initial proof-of-concept dashboard built with Streamlit. Kept for historical/development reference. |
| models | The central "model registry" containing the final, serialized model files ready for deployment. |
| notebooks | The Jupyter Notebooks detailing the R&D, training, and evaluation of all four AI models. |
| reports | (Start Here) Comprehensive project, model performance, and EDA reports. |
| visuals | A repository of charts and plots generated during the data analysis and model evaluation phases. |
To run the entire platform locally, you will need to set up the backend and frontend separately.
- Git
- Conda / Python 3.9+
- Node.js 16+
- Google Gemini API Key (Free tier is sufficient)
The models and notebooks require training data which is not included in this repo.
➡️ Follow the instructions in the data/README.md to download the necessary datasets. The trained models are located in the models/ directory.
The backend API serves the models from the models directory.
➡️ Follow the setup instructions in the disaster-insight-api/README.md to run the FastAPI server.
The frontend is the user-facing application.
➡️ Follow the setup instructions in the disaster-insight-frontend/README.md to run the React app.
Muhammad Zain
Data Scientist | AI Engineer | Applied ML Developer | LLM Developer
The AI models in this project were trained on several publicly available datasets. We are grateful to the creators and maintainers of these resources.
➡️ For a complete list of datasets, sources, and their respective licenses, please see the data/README.md.
The source code for this project is licensed under the Apache License 2.0.
Please see the LICENSE file for the full text. This permissive license allows for commercial and non-commercial use, modification, and distribution.
The datasets used to train the models are subject to their own original licenses, some of which are non-commercial. Please refer to the data/README.md for detailed information on data sources and their respective terms of use before using them for any purpose.