🚨 Text Based Emergency Detection AI

An end-to-end AI application that detects whether a text message indicates an emergency.

This project combines Natural Language Processing (NLP), a deep learning model, and a full-stack system with database and deployment.

---

📌 Overview

Users can:

• Enter a message
• Get an AI prediction (Emergency / Not Emergency)
• View message history stored in a database

The entire system is fully containerized and can be run locally with a single command.

---

🧠 How It Works (Step-by-Step)

1. User enters a message in the frontend (Next.js UI)

2. Frontend sends request to backend: POST /predict

3. Backend processes text:

• Cleans text (lowercase, remove symbols, URLs, etc.)
• Tokenizes into words
• Converts words to numerical indices using a vocabulary
• Pads or truncates to fixed length

4. Model prediction:

• Bidirectional LSTM processes the sequence
• Outputs logits → converted into prediction (emergency / not emergency)

5. Database storage:

• Message and prediction are saved into PostgreSQL

6. Response:

• Result is returned and displayed to user

7. History retrieval: GET /history

• Fetches latest messages from database

---

🏗️ Tech Stack

Backend:

• FastAPI
• PyTorch (BiLSTM model)
• psycopg2

Frontend:

• Next.js (React)
• TailwindCSS

Database:

• PostgreSQL

DevOps / Deployment:

• Docker
• Docker Compose
• AWS EC2
• GitHub Actions (CI/CD)

---

📂 Project Structure

disaster_classification/ │ ├── backend/ │ ├── src/ │ │ ├── api.py │ │ ├── model.py │ │ ├── inference.py │ │ └── dataset.py │ ├── Dockerfile │ └── requirements.txt │ ├── frontend/ │ ├── app/ │ ├── Dockerfile │ └── package.json │ ├── docker-compose.yml └── README.md

---

🚀 Getting Started (Local Setup)

Prerequisites

• Docker
• Docker Compose

---

1. Clone the Repository

git clone https://github.com/moecrosoft/disaster_classification.git cd disaster_classification

---

2. Run the Entire System

docker-compose up -d --build

This will start:

• Frontend → http://localhost:3000
• Backend → http://localhost:8000
• PostgreSQL database

---

3. Open the Application

http://localhost:3000

---

4. Stop the Application

docker-compose down

---

📊 Model Details

• Architecture: Bidirectional LSTM
• Embedding size: 128
• Hidden size: 128
• Sequence length: 60
• Output: 2 classes (Emergency / Not Emergency)

Training Features:

• Text cleaning and normalization
• Vocabulary with and
• Rare word filtering (min_freq)
• Class imbalance handling:
  • Weighted loss
  • WeightedRandomSampler

Evaluation Metrics:

• Accuracy
• F1 Score (~0.72 on validation set)

---

🎯 Features

• Real-time AI prediction
• Message history tracking
• Color-coded results:
  • 🔴 Emergency
  • 🟢 Not Emergency
• Fully containerized system
• One-command local deployment

---

🌐 Deployment

This project is deployed on AWS EC2 using:

• Docker Compose for container orchestration
• GitHub Actions for automatic deployment on push to main

Deployment workflow:

1. Push to GitHub
2. GitHub Actions SSH into EC2
3. Pull latest code
4. Build Docker containers
5. Restart services

---

🔗 API Endpoints

POST /predict

Request: { "text": "Someone collapsed and needs help" }

Response: { "prediction": "emergency (confidence: 0.91)" }

---

GET /history

Response: [ { "user_message": "...", "ai_result": "..." } ]