A Next.js web application for extracting structured cancer information from clinical text using AI-powered analysis.

🌐 Live Demo: https://medical-extraction.vercel.app

Note: The frontend is always live, but the backend EC2 server may not be running at all times to save costs (this is an experimental project). If you encounter connection errors, please reach out.

This is the frontend interface to serve a fine-tuned Llama 3.1 8B model for medical information extraction. Built with Next.js and deployed on Vercel, it provides a user-friendly web interface to extract structured cancer information from clinical text using LLM-powered analysis.

Given clinical text, the model extracts 7 structured fields:

graph LR
    A[User Browser<br/>React Frontend] -->|HTTPS| B[Vercel Edge<br/>Next.js API Routes]
    B -->|HTTP| C[EC2 Backend<br/>FastAPI + vLLM]
    C -->|JSON Response| B
    B -->|JSON Response| A

Data Flow:

1. User enters clinical text in the web form
2. Frontend calls /api/extract (Next.js API route)
3. API route proxies request to EC2 FastAPI gateway (port 8080)
4. FastAPI forwards to vLLM for inference (~2-3 seconds)
5. Structured JSON response displayed in UI

✨ Real-time Extraction - Extract medical entities in 2-3 seconds
🔒 Secure Architecture - EC2 backend IP hidden via server-side proxy
📱 Responsive Design - Works seamlessly on mobile and desktop
🎯 Type-safe - Full TypeScript coverage with strict mode
⚡ Fast & Modern - Built with Next.js 16 and TailwindCSS v4
🔄 Auto-deploy - Push to main → live on Vercel instantly

1. Enter Clinical Text - Paste any cancer-related medical notes
2. AI Processing - Fine-tuned Llama 3.1 8B analyzes the text
3. Structured Output - Get 7 organized fields in seconds
4. Review Results - Cancer type, stage, mutations, treatments, and more

Input:

Patient diagnosed with stage 3 breast cancer with HER2 positive marker. 
Underwent mastectomy followed by adjuvant chemotherapy with trastuzumab 
and pertuzumab. Post-treatment scans show complete response.

Output:

• Cancer Type: Breast cancer
• Stage: 3
• Biomarker: HER2 positive
• Treatment: Mastectomy; trastuzumab and pertuzumab
• Response: Complete response
• (Other fields: null if not mentioned)

Try it yourself: https://medical-extraction.vercel.app

When non-medical text is provided, the system correctly returns flagged message.

src/
├── app/
│   ├── api/
│   │   ├── extract/route.ts   # Proxy to backend /api/v1/extract
│   │   └── health/route.ts    # Proxy to backend /health
│   ├── layout.tsx             # Root layout with metadata
│   └── page.tsx               # Main extraction form page
├── components/
│   ├── ui/                    # ShadcnUI components
│   ├── extraction-form.tsx    # Clinical text input form
│   └── extraction-results.tsx # Results display grid
├── lib/
│   ├── api.ts                 # API client functions
│   └── utils.ts               # Utility functions
└── types/
    ├── api.ts                 # TypeScript interfaces
    └── index.ts               # Barrel exports

docs/
├── SETUP.md                   # Detailed setup guide
├── EXAMPLES.md                # Sample clinical texts for testing
└── BACKEND-REPO-INSTRUCTIONS.md  # Backend context

Want to run this locally or contribute? See docs/SETUP.md for:

• Local development setup
• Environment configuration
• Deployment instructions
• Troubleshooting guide

This is Stage 3 of a 4-stage medical LLM serving project:

The system uses a fine-tuned Llama 3.1 8B model (qLoRA 4-bit quantization) for structured medical entity extraction from clinical text.

Note: BACKEND_API_URL is a server-only variable (no NEXT_PUBLIC_ prefix), meaning it's never exposed to the browser and can be changed in Vercel without redeploying.

This project is deployed on Vercel with auto-deployment from the main branch.

Deployed URL: https://medical-extraction.vercel.app

For deployment details, see docs/SETUP.md.

// POST /api/extract
{
  "text": "Clinical text here...",
  "temperature": 0.3,    // optional
  "max_tokens": 512      // optional
}

// Response
{
  "cancer_type": "melanoma",
  "stage": "IV",
  "gene_mutation": null,
  "biomarker": "PD-L1 5%",
  "treatment": "nivolumab",
  "response": "partial response",
  "metastasis_site": "brain",
  "raw_output": "...",
  "tokens_used": 116
}

• Setup Guide - Complete installation and deployment steps
• Clinical Examples - Sample texts for testing
• Backend Instructions - Backend repo context

• Backend Repository - vLLM server + FastAPI gateway (Stage 1 & 2)
• Fine-tuned Model - loghoag/llama-3.1-8b-medical-ie
• Base Model - meta-llama/Llama-3.1-8B

MIT License - see LICENSE file for details.