Curalink - AI Medical Research Assistant

A full-stack MERN application that acts as a health research companion. It understands user context, retrieves high-quality medical research from multiple sources, reasons over it with an open-source LLM (Llama 3.3-70B via Groq), and delivers structured, source-backed answers...

---

Architecture & AI Pipeline

User Query + Patient Context
         │
         ▼
┌─────────────────────────┐
│ 1. Query Expansion (LLM)│  → "deep brain stimulation" + "Parkinson's"
│    Llama 3.3-70B / Groq │    → "deep brain stimulation Parkinson's disease neurological outcomes"
└─────────┬───────────────┘
          │ expanded query
          ▼
┌───────────────────────────────────────────────────────┐
│ 2. Parallel Retrieval (50–300 results)                 │
│   ┌──────────────┐  ┌──────────────┐  ┌────────────┐ │
│   │  PubMed      │  │  OpenAlex    │  │ ClinTrials │ │
│   │  80 articles │  │  100 articles│  │ 50 trials  │ │
│   └──────────────┘  └──────────────┘  └────────────┘ │
└───────────────────────────────────────┬───────────────┘
                                        │ ~230 raw results
                                        ▼
┌─────────────────────────────────────────────────────────┐
│ 3. Re-Ranking + Deduplication                            │
│   Score = Keyword Relevance (40) + Recency (30)          │
│          + Source Credibility (20) + Abstract Quality (10)│
│   → Top 7 publications + Top 5 clinical trials           │
└───────────────────────────────────────┬─────────────────┘
                                        │ 12 top results
                                        ▼
┌─────────────────────────────────────────┐
│ 4. LLM Reasoning (Llama 3.3-70B/Groq)  │
│   System prompt = patient context       │
│                 + top publications      │
│                 + top trials            │
│                 + conversation history  │
│   Output: structured JSON response      │
└───────────────────────────────────────-┘
                    │
                    ▼
    Structured Response:
    ├── Condition Overview
    ├── Research Insights (cited [PUB1], [PUB2]...)
    ├── Source Publications (cards with URL)
    ├── Clinical Trials (status, eligibility, location)
    ├── Personalized Note
    └── Key Takeaways

---

Tech Stack

Layer	Technology
Frontend	React + Vite
Backend	Express.js + Node.js
Database	MongoDB Atlas
LLM	Groq API - llama-3.3-70b-versatile (open-source model)
Publications	OpenAlex API + PubMed NCBI Entrez API
Clinical Trials	ClinicalTrials.gov API v2
Frontend Deploy	Vercel
Backend Deploy	Render

Why Groq + Llama 3.3-70B? Groq provides free, ultra-fast cloud inference for open-source model weights (Llama 3.3 by Meta). This satisfies the "no Gemini/OpenAI" requirement while delivering best-in-class reasoning quality.

---

Project Structure

curalink/
├── backend/
│   ├── server.js                    # Express app entry point
│   ├── .env.example                 # Environment variables template
│   ├── render.yaml                  # Render.com deploy config
│   ├── models/
│   │   └── Session.js               # MongoDB session + message schema
│   ├── routes/
│   │   ├── chat.js                  # POST /api/chat (main pipeline)
│   │   └── sessions.js              # GET/POST/DELETE /api/sessions
│   └── services/
│       ├── llmService.js            # Groq API (Llama 3.3-70B)
│       ├── queryExpander.js         # LLM-based query expansion
│       ├── pubmedService.js         # PubMed esearch + efetch pipeline
│       ├── openAlexService.js       # OpenAlex works API
│       ├── clinicalTrialsService.js # ClinicalTrials.gov v2 API
│       ├── ranker.js                # Multi-factor re-ranking engine
│       └── pipelineService.js       # Pipeline orchestrator
│
└── frontend/
    ├── index.html
    ├── vite.config.js
    ├── vercel.json                  # Vercel deploy config
    └── src/
        ├── App.jsx                  # Main layout + session state
        ├── api.js                   # Axios API client
        ├── index.css                # Design system (dark glassmorphism)
        └── components/
            ├── SessionSidebar.jsx   # Session history sidebar
            ├── InputPanel.jsx       # Structured + quick chat input
            ├── MessageCard.jsx      # User bubble + AI response card
            ├── PublicationCard.jsx  # Individual publication card
            └── TrialCard.jsx        # Individual clinical trial card

---

Local Setup

Prerequisites

• Node.js 18+
• A free MongoDB Atlas cluster
• A free Groq API key (Llama 3.3-70B)

1. Clone & Install

# Install backend dependencies
cd curalink/backend
npm install

# Install frontend dependencies
cd ../frontend
npm install

2. Configure Environment Variables

Backend (curalink/backend/.env):

MONGODB_URI=mongodb+srv://<user>:<pass>@<cluster>.mongodb.net/curalink
GROQ_API_KEY=gsk_xxxxxxxxxxxxxxxxxxxxxxxxxxxx
FRONTEND_URL=http://localhost:5173
PORT=5000

Frontend (curalink/frontend/.env):

VITE_API_URL=http://localhost:5000/api

3. Start Development Servers

# Terminal 1: Backend
cd curalink/backend
npm run dev

# Terminal 2: Frontend
cd curalink/frontend
npm run dev

Open http://localhost:5173

---

Deployment

Backend → Render.com

1. Push the backend/ folder to GitHub
2. Create a new Web Service on render.com
3. Set build command: npm install
4. Set start command: npm start
5. Add environment variables: MONGODB_URI, GROQ_API_KEY, FRONTEND_URL

Frontend → Vercel

1. Push the frontend/ folder to GitHub
2. Import project on vercel.com
3. Set environment variable: VITE_API_URL=https://your-render-url.onrender.com/api
4. Deploy — Vercel auto-detects Vite

---

Example Queries to Demo

Query	Disease	Notes
Latest treatment for lung cancer	Lung Cancer	Tests publication retrieval
Clinical trials for diabetes	Diabetes	Tests ClinicalTrials.gov integration
Deep Brain Stimulation outcomes	Parkinson's disease	Tests query expansion
Top researchers in Alzheimer's	Alzheimer's disease	Tests OpenAlex author data
Can I take Vitamin D?	Lung Cancer	Tests multi-turn context

---

API Endpoints

POST /api/chat

Main research pipeline endpoint.

Request body:

{
  "message": "Latest treatment for lung cancer",
  "sessionId": "optional-existing-session-id",
  "patientName": "John Smith",
  "disease": "Lung Cancer",
  "location": "Toronto, Canada"
}

Response:

{
  "sessionId": "uuid",
  "result": {
    "expandedQuery": "lung cancer treatment immunotherapy targeted therapy",
    "llmResponse": {
      "conditionOverview": "...",
      "researchInsights": [...],
      "clinicalTrialsSummary": "...",
      "personalizedNote": "...",
      "keyTakeaways": [...]
    },
    "publications": [...],
    "clinicalTrials": [...],
    "stats": {
      "totalRetrieved": 230,
      "pubmedCount": 80,
      "openAlexCount": 100,
      "clinicalTrialsCount": 50
    }
  }
}

GET /api/sessions — List all sessions

GET /api/sessions/:id — Get session with history

POST /api/sessions — Create new session

DELETE /api/sessions/:id — Delete session

GET /health — Health check

---

Design Decisions

LLM Choice: Groq (Llama 3.3-70B)

• Open-source model: Llama 3.3 weights are fully open (Meta license)
• Not OpenAI/Gemini: Satisfies hackathon prohibition
• Speed: Groq's LPU delivers ~800 tokens/sec — no cold starts
• Quality: 70B parameter model excels at medical reasoning and JSON output

Retrieval Strategy: Hybrid depth-first

• Retrieves broad candidate pool (230+ results) before filtering
• PubMed prioritized for credibility (score: 20 vs OpenAlex: 15)
• Deduplication by normalized title to avoid showing same paper twice

Re-ranking: Multi-factor scoring (no vector DB needed)

• TF-style keyword matching (title: 4x weight, abstract: 1x)
• Recency decay: linear falloff over 10 years
• Fast, deterministic, transparent — no embedding API needed

Multi-turn Context

• Last 6 messages injected into LLM system prompt
• Session stored in MongoDB with full metadata per message
• Patient context (disease, name, location) persisted across turns