Last summer, when I found out about the Gemini API Developer Competition, I saw it as a great chance to get my hands dirty with GenAI applications. As fitness enthusiasts, we (me & Manos Chainakis) thought of creating an app that could generate personalised workout and nutrition plans—combining AI with the preferences of human coaches. That’s how Fitness Tribe AI was born. This post will walk you through the development process and the tech stack I used, with focus on the GenAI aspect.

Fitness Tribe AI  |  Gemini API Developer Competition  |  Google AI for Developers

Making Personal Training Better for Everyone, Everywhere.

ai.google.dev

The Concept Behind Fitness Tribe AI

Fitness Tribe AI combines the expertise of human coaches with the capabilities of AI models to create custom fitness programs that meet each athlete's needs and goals.

The Tech Stack

The main components of the tech stack are:

• FastAPI for the backend and AI model integration
• Supabase for user authentication and data management
• Ionic & Angular for the frontend mobile app
• Astro for the landing page

FastAPI: Backend and AI Integration

FastAPI serves as the backbone of Fitness Tribe AI, handling the AI-powered analysis.

Here’s how the project is structured:

fitness-tribe-ai/ ├── app/ │   ├── main.py              # Entry point for FastAPI app │   ├── routers/             # Handles API routes (meals, nutrition, workouts) │   ├── models/              # Manages interactions with AI models │   ├── schemas/             # Pydantic models for input validation │   ├── services/            # Business logic for each feature 

Key elements of the FastAPI implementation:

• API Routing: Routes are divided into separate files for meals (meals.py), workouts (workouts.py), and nutrition (nutrition.py), keeping the API structure organised and scalable. Each router is connected in main.py, where FastAPI’s routing system ties everything together.

from fastapi import FastAPI from app.routers import meals, nutrition, workouts  app = FastAPI()  app.include_router(meals.router) app.include_router(nutrition.router) app.include_router(workouts.router) 

• Gemini Model Integration: The GeminiModel class, in gemini_model.py, handles the AI model interaction. Taking as an example the meal analyzer method, I am using Pillow to process image data, and the app sends both the image and a custom prompt to the Gemini AI to analyze meal details. An important detail here is that the prompt should be specific enough, when it comes to the format of the expected response, so that it can be processed by the service layer.

class GeminiModel:     @staticmethod     def analyze_meal(image_data):         prompt = (             "Analyze the following meal image and provide the name of the food, "             "total calorie count, and calories per ingredient..."             "Respond in the following JSON format:"             "{'food_name': '<food name>' ...}"         )         image = Image.open(BytesIO(image_data))         response = model.generate_content([prompt, image])         return response.text 

• Pydantic Schema for Data Validation: The response from the AI model is validated and structured using Pydantic models. For instance, the Meal schema in schemas/meal.py ensures the response is consistent before it is returned to the user.

from pydantic import BaseModel from typing import Dict  class Meal(BaseModel):     food_name: str     total_calories: int     calories_per_ingredient: Dict[str, int] 

• Service Layer: The service layer, located in services/, encapsulates the logic of each feature. For example, the meal_service.py handles the meal analysis, ensuring the data is properly processed before returning the AI results.

from app.models.gemini_model import GeminiModel from app.schemas.meal import Meal from fastapi import HTTPException import logging import json  def analyze_meal(image_data: bytes) -> Meal:     try:         result_text = GeminiModel.analyze_meal(image_data)         if not result_text:             raise HTTPException(status_code=500, detail="No response from Gemini API")          clean_result_text = result_text.strip("```  json
").strip("  ```")         result = json.loads(clean_result_text)         return Meal(             food_name=result.get("food_name"),             total_calories=result.get("total_calories"),             calories_per_ingredient=result.get("calories_per_ingredient"),         )     except Exception as e:         raise HTTPException(status_code=500, detail=str(e)) 

By leveraging FastAPI’s modular structure, clear API routing, Pydantic for data validation, and well-organized service logic, Fitness Tribe AI efficiently handles AI model interactions with custom prompts to deliver personalized fitness and nutrition insights. You can find the full repo here:

fitness-tribe-ai/ ├── app/ │   ├── __init__.py │   ├── main.py │   ├── models/ │   │   ├── __init__.py │   │   ├── gemini_model.py │   ├── routers/ │   │   ├── __init__.py │   │   ├── meals.py │   │   ├── nutrition.py │   │   ├── workouts.py │   ├── schemas/ │   │   ├── __init__.py │   │   ├── meal.py │   │   ├── nutrition.py │   │   ├──

Supabase: User Management & Auth

For user authentication and account management, I used Supabase, which provided a secure, scalable solution without requiring a custom-built authentication system.

Key features I leveraged:

• Authentication: Supabase's built-in authentication enabled users to log in and manage their profiles with ease.

• Database Management: Using Supabase’s PostgreSQL-backed database, I stored user preferences, workout routines, and meal plans to ensure updates reflected immediately in the app.

Ionic & Angular: Cross-Platform Frontend

For the frontend, I chose Ionic and Angular, which enabled me to create a mobile-first app that could be deployed on the web right away while it could also be shipped as native for both iOS and Android.

Astro: A Lightning-Fast Landing Page

For the landing page, I opted for Astro, which focuses on performance by shipping minimal JavaScript. Astro allowed me to build a fast, lightweight page that efficiently showcased the app.

Conclusion

Developing Fitness Tribe AI was a learning journey that enabled me to explore the power that AI models give us nowadays. Each framework played a role, from FastAPI’s robust backend capabilities and ease of use to Supabase’s user management, Ionic’s cross-platform frontend and Astro’s high-performance landing pages.

For anyone looking to build a GenAI app, I highly recommend exploring these frameworks (and especially FastAPI) for their powerful features and smooth developer experience.

Have questions or want to learn more about it? Let me know in the comments!