🧬 Medical AI Knowledge Agent

A full-stack AI agent that intelligently routes natural language queries between a structured SQL database (patient metrics) and an unstructured vector database (clinical documents). Packaged as decoupled microservices in Docker containers and deployed on AWS EC2.

---

📸 Demo

Live deployment on AWS EC2 (eu-north-1) — natural language input → autonomous tool selection → real-time SQL or RAG response

---

🏗️ System Architecture

+---------------------------------+
|         Next.js Web GUI         |
|      (Port 3000 - Chat UI)      |
+---------------------------------+
                |
                | HTTP POST /chat (User Query)
                v
+---------------------------------+
|         FastAPI Backend         |
|    (Port 8000 - API Router)     |
+---------------------------------+
                |
                | LangGraph Tool Routing
                v
+---------------------------------+
|       AI Inference Engine       |
|  (Groq LLM + LangGraph Agent)   |
+---------------------------------+
           /              \
     SQL Tool            RAG Tool
          v                 v
+----------------+  +----------------+
| SQLite Patient |  | ChromaDB Vector|
| Database       |  | Index (PDFs)   |
+----------------+  +----------------+

1. Frontend (Next.js): Modern chat interface for interacting with the AI agent in real time.
2. Backend (FastAPI + LangGraph): Asynchronous API that receives queries and uses an LLM to autonomously decide whether to execute a SQL query or a vector similarity search — the agent picks the right tool without being told.
3. Databases: SQLite for structured patient tabular data; ChromaDB for unstructured medical PDF knowledge.
4. Containerisation: Multi-container setup via Docker Compose V2, optimised for resource-constrained cloud environments (AWS EC2 t3.micro).

---

📊 Agent Capabilities

Capability	Detail
Structured queries	SQL via Pandas — patient counts, cholesterol levels, tabular filtering
Unstructured queries	RAG via ChromaDB + HuggingFace embeddings — clinical definitions, medical guidelines from PDFs
Autonomous routing	LangGraph agent selects SQL or RAG based on query intent — no manual switching
Session memory	Conversation context persists across turns within a session
Deployment	Live on AWS EC2 — Docker Compose orchestrated, swap-optimised for free tier

---

🛠️ Tech Stack

Layer	Technology
AI & Orchestration	LangChain, LangGraph, Groq API, HuggingFace Embeddings
Databases	ChromaDB (vector), SQLite, Pandas
API	FastAPI, Uvicorn
Frontend	Next.js, React, Node.js, Tailwind CSS
Containerisation	Docker, Docker Compose V2
Cloud	AWS EC2 (Ubuntu Linux, eu-north-1)

---

📁 Repository Structure

medical-knowledge-agent/
│
├── data/                        # Vector DB and SQLite — generated locally
│   ├── chroma_db/               # ChromaDB embeddings (git-ignored)
│   └── patient_db.db            # SQLite patient database (git-ignored)
├── logs/                        # Application and container runtime logs
├── src/
│   ├── agent.py                 # LangGraph tool definitions and routing logic
│   ├── app.py                   # Backend entry point
│   ├── rag_tool.py              # RAG retrieval pipeline
│   ├── rag_test.py              # Unit tests for vector retrieval
│   ├── server.py                # FastAPI endpoints and CORS configuration
│   ├── setup_db.py              # PDF ingestion and SQLite initialisation
│   └── test_db.py               # Database connection validation
├── frontend/
│   ├── public/                  # Static assets
│   └── src/app/
│       ├── page.js              # Main chat UI component
│       ├── layout.js            # Next.js root layout
│       └── globals.css          # Tailwind and global styles
├── Dockerfile.backend           # Docker image — FastAPI backend
├── Dockerfile.frontend          # Docker image — Next.js frontend
├── docker-compose.yml           # Container orchestration
├── requirements.txt             # Python dependencies (CPU-optimised)
└── README.md

Data note: The data/ directory is generated dynamically on the host using setup_db.py and mounted into containers at runtime — keeping the Docker image lean and protecting sensitive data.

---

🚀 How to Run

Prerequisites

• Docker and Docker Compose V2 installed
• A Groq API key (free)

Local Setup

git clone https://github.com/sachinkumarp-code/medical-knowledge-agent
cd medical-knowledge-agent

# Create a .env file with your API key
echo "GROQ_API_KEY=your_key_here" > .env

# Initialise databases locally (run once)
python src/setup_db.py

# Launch the full stack
docker compose up --build -d

Open http://localhost:3000 in your browser.

Example Queries

"How many patients have cholesterol above 250?"      → SQL tool
"What is the definition of cardiac allograft rejection?"  → RAG tool
"Summarise the treatment guidelines for heart failure"    → RAG tool

---

☁️ AWS Deployment — Resource Optimisations

Deployed on EC2 t3.micro (1 vCPU, 1 GB RAM). Three optimisations prevent OOM crashes on free-tier hardware:

1. Virtual RAM — configured a permanent 4 GB /swapfile on the instance
2. CPU-only PyTorch — forced via --extra-index-url https://download.pytorch.org/whl/cpu in requirements.txt, saving over 5 GB compared to the default GPU build
3. Runtime data injection — databases initialised after container startup via docker exec -it medical-knowledge-agent-backend-1 python src/setup_db.py, keeping the Docker image itself lean

---

🔬 Part of a Larger Medical AI Portfolio

Project	Description
Medical AI Knowledge Agent (this repo)	Autonomous RAG + SQL agent with Next.js UI
Brain Tumor MLOps	PyTorch CNN deployed live on AWS EC2 via FastAPI + Streamlit + Docker
🔒 Cardiac Histopathology Segmentation	Attention UNet, active learning, 6,600+ WSIs — ongoing thesis (confidential)

---

👤 Author

Sachinkumar P — M.Tech., Structural & Computational Biology, IIT Roorkee
github.com/sachinkumarp-code · LinkedIn