AI Nutrition Coach

AI Nutrition Coach is a small Flask web app that analyzes a food image and returns a structured nutrition summary. It uses a local multimodal model served by Ollama, so the app can run without external paid APIs.

Overview

The app lets a user:

• upload a food image
• ask a question about that image
• get an estimated calorie and nutrient breakdown
• read a formatted response directly in the browser

The current model backend is gemma3:4b through Ollama's local /api/chat endpoint.

Libraries

This project intentionally keeps the stack small:

• Flask: web server, routing, templates, and flash messages
• requests: HTTP client used to call the local Ollama API
• base64: built-in Python module used to encode uploaded images
• re: built-in Python module used to format model output into HTML
• os: built-in Python module used for environment-based configuration

How It Works

The application flow is organized in the same high-level structure used in app.py:

1. Flask app setup
2. Ollama client and model configuration
3. Uploaded image encoding
4. Response formatting
5. Model inference through chat messages with image input

When a user submits the form, the app:

1. reads the uploaded image
2. converts it to base64
3. sends the prompt, question, and image to Ollama
4. formats the returned text into HTML for display

Requirements

• Python 3.11+
• uv or pip
• Ollama
• the local model gemma3:4b

Setup

Install Python dependencies:

uv sync

Pull the Ollama model:

ollama pull gemma3:4b

Start Ollama:

ollama serve

Run the app:

uv run app.py

Then open:

http://127.0.0.1:5000

Configuration

The app supports these environment variables:

• OLLAMA_URL: overrides the Ollama server URL
• FLASK_SECRET_KEY: sets the Flask secret key

Example:

export OLLAMA_URL=http://127.0.0.1:11434
export FLASK_SECRET_KEY=change-me
uv run app.py

Project Structure

.
├── app.py
├── assets/
│   └── app.png
├── pyproject.toml
├── static/
│   └── style.css
└── templates/
    └── index.html

Notes

• The calorie and nutrient estimates are approximate.
• Response quality depends on the image quality and the local model.
• The app is configured for development and runs with Flask debug mode enabled.