🚨 Disaster VQA Response System

Multimodal AI for Real-Time Disaster Assessment

---

🌍 Overview

The Disaster VQA Response System is a next-generation multimodal AI application that analyzes disaster imagery and answers operational questions in real time.

Instead of simple classification, this system performs context-aware reasoning using Vision-Language Models (VLMs).

Example:

🖼 Image: Flooded highway
❓ Question: Is the road passable?
🤖 Answer: No, water levels appear unsafe for vehicle passage.
📊 Confidence: 87%

🧠 System Architecture

Image → Vision Encoder → Image Embeddings  
Question → Language Encoder → Text Embeddings  
Cross-Modal Fusion → Contextual Reasoning  
Output → Answer + Confidence Score

---

⚙️ Tech Stack

🔹 AI / ML Model Specifications

Active Model: Salesforce/blip-vqa-base

While the initial design accounted for heavy-weight models like LLaVA and Qwen-VL, the current active implementation utilizes BLIP (Bootstrapping Language-Image Pre-training).

Why BLIP?

• Hardware Efficiency: BLIP is a highly capable VLM that requires significantly less VRAM (~1GB-2GB) compared to the massive 16GB+ requirements of LLaVA 7B. This allows the API to run smoothly on edge devices and consumer GPUs without out-of-memory (OOM) crashes.
• Speed: Inference time is dramatically reduced, enabling true real-time disaster reasoning.

(See the guide below if you wish to run the full LLaVA architecture).

🔹 Technologies

• PyTorch
• Hugging Face Transformers
• FastAPI (Backend API)
• Python

🔹 Frontend

• AI-assisted UI (Antigravity + Gemini Pro AI)
• Futuristic dashboard

🔹 Deployment

• Docker
• GPU-enabled Cloud Hosting

---

🚀 Features

• 📷 Image Upload
• 💬 Natural Language Questions
• 🤖 Contextual AI Responses
• 📊 Confidence Score
• ⚡ Real-Time Inference
• 📜 Session History

---

📂 Project Structure

Disaster-VQA-Response-System/
│
├── backend/
│   ├── main.py (FastAPI Server)
│   ├── ml/model_pipeline.py (VQA Inference)
│   └── routers/
├── frontend/
│   ├── index.html
│   ├── style.css
│   └── script.js
├── models/
├── docker/
└── README.md

---

⚡ Development Note

The core logic, model integration, backend architecture, and AI inference pipeline are fully self-designed and implemented using Python, FastAPI, and PyTorch.

The frontend interface was developed with the assistance of AI-powered tools (Antigravity and Gemini Pro AI) to accelerate modern UI development, while maintaining full backend ownership and integration control.

---

🛠️ Running with LLaVA 1.5 (Advanced Setup)

If you have access to a High-End GPU (e.g., RTX 3090, 4090, or A100 with 16GB+ VRAM) and wish to utilize the LLaVA 1.5 7B model instead of BLIP, follow these steps to modify the backend:

1. Install Quantization Libraries
   To load LLaVA efficiently, assure bitsandbytes is installed to support 4-bit load:

   pip install bitsandbytes accelerate

2. Modify the Pipeline Code
   Open backend/ml/model_pipeline.py and replace the BLIP classes with LLaVA:

   # Replace these imports:
   # from transformers import BlipProcessor, BlipForQuestionAnswering
   
   # With LLaVA imports:
   from transformers import AutoProcessor, LlavaForConditionalGeneration
   from transformers import BitsAndBytesConfig

3. Update Model Initialization
   Change the __init__ and load_model methods:

   def __init__(self, model_id="llava-hf/llava-1.5-7b-hf"):
       # ... setup code ...
       
   def load_model(self):
       self.processor = AutoProcessor.from_pretrained(self.model_id)
       
       bnb_config = BitsAndBytesConfig(
           load_in_4bit=True,
           bnb_4bit_compute_dtype=torch.float16
       )
       
       self.model = LlavaForConditionalGeneration.from_pretrained(
           self.model_id,
           quantization_config=bnb_config,
           low_cpu_mem_usage=True
       )

4. Update Inference Prompt
   LLaVA requires a specific prompt template format. Update the generate pipeline:

   prompt = f"USER: <image>\n{question}\nASSISTANT:"
   inputs = self.processor(text=prompt, images=image, return_tensors="pt")

Restart the FastAPI server. Note: The initial start will download roughly 10GB of weights from Hugging Face.

---

🌍 Future Scope

• 🎥 Video-based Disaster Analysis
• 🗺 GIS Mapping Integration
• 🚨 Automated Alert Systems
• 📡 Edge Deployment
• 🌐 Multilingual Support

---

👨‍💻 Author

Gautam N Chipkar
B.E – Artificial Intelligence & Data Science

---

⭐ Support

If you find this project valuable, consider giving it a star ⭐