HealthifyAI — Food Product Analyzer

ML Engineers: The LLM finetuning, dataset preparation, and model training code lives here →
github.com/adarshkatare6/HeathifyAI-ML

---

What is HealthifyAI?

HealthifyAI is a full stack web application where users upload an image of a food ingrediants, the system extracts text via OCR and returns a simple, clear health review powered by a finetuned LLM — all behind a secure JWT-authenticated REST API.

---

Live Demo

Service	Platform	URL
Frontend	Vercel	https://healthify-ai-zeta.vercel.app/
Auth + API Server	Render
ML Inference API	HuggingFace + ngrok
Model	HuggingFace Hub	adarsh6/healthifyai-base-lora-merged

---

Architecture

User visits Vercel (React frontend)
        ↓
Login / Register page
        ↓
POST /login or /register → Render Auth Server
        ↓
Flask verifies credentials against MongoDB Atlas
        ↓
Returns JWT token (24hr expiry) to React
        ↓
React stores token in memory (not localStorage)
        ↓
User uploads image → React converts to base64
        ↓
POST /predict + Authorization: Bearer <token> → Render
        ↓
Render validates JWT → forwards image to Colab via ngrok
        ↓
Colab: OCR (Tesseract) → text cleaning → LLM inference
        ↓
Response returned → Render → React → displayed to user

---

Project Structure

HealthifyAI/
├── frontend/                  → Deployed on Vercel
│   ├── src/
│   │   ├── App.jsx            → Main app, token state, conditional routing
│   │   ├── Login.jsx          → Login page with cold-start spinner
│   │   ├── Register.jsx       → Register page
│   │   └── main.jsx
│   ├── index.html
│   ├── package.json
│   ├── vite.config.js
│   ├── vercel.json
│   └── .env.example           
│
└── backend/                   → Deployed on Render
    ├── app.py                 → Flask auth + proxy server
    ├── requirements.txt
    └── render.yaml

---

---

Security

• Passwords hashed with bcrypt before storing in MongoDB — never stored in plaintext
• JWT used for stateless authentication with 24-hour expiry
• Colab ngrok URL is only known to Render — never exposed to frontend
• JWT validated on every /predict call as a backend security layer
• Token stored in React state (memory), not localStorage — eliminates XSS-based token theft
• 401 Unauthorized response automatically clears token and redirects to login
• All sensitive config (MONGO_URI, JWT_SECRET, COLAB_URL) stored as environment variables — never hardcoded

---

---

How to Run Locally

Backend:

cd backend
pip install -r requirements.txt
export MONGO_URI=your_mongo_uri
export JWT_SECRET=your_secret
export COLAB_URL
python app.py

Frontend:

cd frontend
cp .env.example .env
# Set VITE_AUTH_SERVER_URL=http://localhost:5000
npm install
npm run dev

Visit http://localhost:5173

---

Starting the ML Inference Backend (Colab)

Do this before testing predictions.

1. Open the Colab notebook from the ML repo
2. Runtime → Change runtime type → T4 GPU
3. Runtime → Run All
4. Connect to ngrok
5. Wait ~3-4 minutes for model to load
6. Backend is live

---

ML Side of the Project

All model training code, dataset preparation, LoRA finetuning, and evaluation is in a separate repo:

github.com/adarshkatare6/HeathifyAI-ML

• Base model: Mistral-7B-v0.3
• Finetuning: LoRA (merged into base)
• Quantization: 4-bit (nf4, double quant, float16 compute) for T4 VRAM
• Hosted on HuggingFace: adarsh6/healthifyai-base-lora-merged