Doctor's Aide (MedHelp)

AI-assisted patient intake app with voice capture, structured extraction, and offline persistence.

Features

• 🎙️ Voice-to-text via browser Web Speech API (no paid STT needed)
• 🤖 AI data extraction using OpenRouter's Xiaomi MiMo-V2-Flash model
• 💾 Offline persistence - Patient records saved to browser localStorage
• 🔐 Google sign-in via Firebase Authentication
• 🔍 Search & Filter - Filter patients by name, symptoms, or date (dd/mm/yy)
• 🧪 Dev Mode - Generate random test patients with AI extraction (hidden in production)
• ⚡ Modern Stack - Vite + React + TypeScript + Tailwind + shadcn/ui

Why Xiaomi MiMo-V2-Flash?

We chose Xiaomi MiMo-V2-Flash as our AI extraction model for several compelling reasons:

Model Specifications

• Architecture: Mixture-of-Experts (MoE) with 309B total parameters, 15B active
• Context Window: 262,144 tokens (256K)
• Cost: FREE via OpenRouter ($0/M input & output tokens)

Rankings & Performance

• 🏆 #1 Open-Source Model on SWE-bench Verified and SWE-bench Multilingual
• 📊 Comparable to Claude Sonnet 4.5 at only ~3.5% of the cost
• ⚡ High Throughput: ~69 tokens/second average
• 🟢 100% Uptime on OpenRouter

Why It Works for Healthcare

• Excels at reasoning and structured extraction - perfect for parsing clinical notes
• Supports hybrid-thinking mode for complex medical inference
• Open-source with transparent model weights
• Zero cost makes it ideal for prototyping and small clinics

Previous Models Tried

We evaluated several models before settling on MiMo-V2-Flash:

• Groq llama-3.3-70b-versatile - Required paid API key
• Google Gemini - API quota issues
• NVIDIA Nemotron 3 Nano 30B - Did not follow JSON output instructions
• DeepSeek R1 - Also failed to return structured JSON

MiMo-V2-Flash provides the best balance of cost (free), capability (structured JSON extraction), and reliability.

Deployment: Netlify → Vercel Migration

We migrated from Netlify to Vercel for the following reasons:

Why We Switched

1. Better SPA Routing - Vercel handles client-side routing natively with vercel.json
2. Faster Builds - Vercel's build times are noticeably faster for Vite projects
3. Better DX - Automatic preview deployments and instant rollbacks
4. Edge Functions - Better support for serverless functions if needed later

Vercel Configuration

The project includes a vercel.json file that handles SPA routing:

{
  "rewrites": [
    { "source": "/(.*)", "destination": "/" }
  ]
}

This ensures that refreshing on any route (e.g., /dashboard) doesn't result in a 404.

Quick Start

git clone https://github.com/vijayho1/MedHelp.git
cd MedHelp
npm install
cp .env.example .env   # create this file using the template below
npm run dev

Open http://localhost:5173 and sign in with Google, then add a patient (voice or typed). Data persists in your browser.

Environment Variables (.env)

# OpenRouter API (FREE - for AI extraction)
VITE_OPENROUTER_API_KEY=your_openrouter_key

# Firebase Auth
VITE_FIREBASE_API_KEY=AIzaSyD-iw-DFQQBarKfVbc0j1FsVcX2JMj2qB8
VITE_FIREBASE_AUTH_DOMAIN=medhelp-af68a.firebaseapp.com
VITE_FIREBASE_PROJECT_ID=medhelp-af68a
VITE_FIREBASE_APP_ID=1:999268392916:web:71289a0c5fb6725d1a8041
VITE_FIREBASE_STORAGE_BUCKET=medhelp-af68a.firebasestorage.app
VITE_FIREBASE_MESSAGING_SENDER_ID=999268392916

Getting an OpenRouter API Key (Free)

1. Go to OpenRouter.ai
2. Sign up / Sign in
3. Navigate to API Keys
4. Create a new key and add it to your .env

Note: MiMo-V2-Flash is completely free to use!

Deployment (Vercel)

1. Push your code to GitHub
2. Import project in Vercel Dashboard
3. Add environment variables in Project Settings → Environment Variables
4. Deploy! Vercel auto-deploys on every push to main

Firebase Auth Setup

Ensure your Vercel domain is whitelisted:

1. Go to Firebase Console
2. Navigate to Authentication → Settings → Authorized domains
3. Add your .vercel.app domain

Tech Stack

Category	Technology
Framework	Vite + React 18
Language	TypeScript
Styling	Tailwind CSS + shadcn/ui
AI	OpenRouter (MiMo-V2-Flash)
Auth	Firebase Authentication
Speech	Browser Web Speech API
Storage	localStorage
Hosting	Vercel

Project Structure

src/
├── components/
│   ├── PatientForm.tsx      # Main form with AI extraction
│   ├── PatientCard.tsx      # Patient display card
│   ├── VoiceRecorderWebSpeech.tsx  # Voice recording
│   └── ui/                  # shadcn/ui components
├── contexts/
│   ├── AuthContext.tsx      # Firebase auth state
│   └── PatientContext.tsx   # Patient CRUD + random generation
├── pages/
│   ├── Dashboard.tsx        # Main dashboard with filters
│   ├── Login.tsx            # Google sign-in
│   └── Index.tsx            # Landing page
└── types/
    └── patient.ts           # TypeScript interfaces

Features in Detail

Voice Recording

• Uses browser's native Web Speech API
• No external STT service needed
• Works offline (for transcription)

AI Extraction

The AI extracts structured data from clinical notes:

• Age, History, Symptoms
• Tests, Allergies
• Possible Condition (AI-inferred diagnosis)
• Recommendations (follow-up suggestions)

Date Filtering

Filter patients by creation date using dd/mm/yy or dd/mm/yyyy format.

Dev Mode Features

In development (npm run dev), a "Generate Random Patients" button appears to create test data with AI extraction. This is hidden in production builds.

Troubleshooting

Issue	Solution
AI errors	Verify VITE_OPENROUTER_API_KEY is set
Auth blocked	Add Vercel domain to Firebase authorized domains
404 on refresh	Ensure vercel.json is in project root
Duplicate transcription	Fixed in latest update - pull latest changes
No patients after refresh	localStorage was cleared; data is per-device

Future Improvements

• Replace localStorage with cloud database (Supabase/Firestore)
• Add audit logs per user
• Role-based access for clinics
• Export patient data (PDF/CSV)
• Multi-language support for voice input

License

MIT

---

Built with ❤️ using React, TypeScript, and AI