How I Built a GenAI Workout Coach Using JSON + Few-shot Prompting

Harshada Javeri — Mon, 21 Apr 2025 06:31:19 +0000

Why This Project?
As someone passionate about fitness and fascinated by the power of Generative AI, I set out to build something that could benefit both gym regulars and beginners: a personalized workout and diet coach powered by GenAI.

Many people struggle with crafting a fitness routine tailored to their body, schedule, and goals. I wanted to build an AI system that could take a few simple preferences and return actionable, structured plans—in seconds.

Thus, FitFusion was born!

💡 What It Does
FitFusion is an AI-powered tool that generates:

💪 Personalized workout routines

🍎 Balanced meal plans based on fitness goals

🕒 Time-sensitive recommendations (quick vs. long sessions)

Users provide simple inputs like:

Fitness goal (e.g., weight loss, muscle gain)

Time per day

Injuries (if any)

Equipment available

Diet preference

The AI returns a JSON-formatted plan that includes:

Daily workouts (with sets, reps, rest)

Meal suggestions

Tips or cautions

🧰 GenAI Capabilities Used
Here are the three main GenAI techniques used in this project:

✅ JSON Mode / Structured Output
I used the model's structured output mode to format the workout and meal plans in clean, machine-readable JSON. This makes the output easier to consume in an app, chatbot, or even a mobile UI.
✍️ Few-shot Prompting
To guide the AI on how to respond, I used few-shot examples. A few sample user profiles and their ideal plans helped the model learn the expected structure, tone, and content.
🔍 Grounding + Context Control (Optional Add-On)
To reduce hallucinations, I added brief scientific context snippets (like resting periods, macro distribution), which I plan to expand into retrieval augmented generation (RAG) in future versions.

⚖️ Trade-offs & Challenges
Like any GenAI project, this one wasn’t perfect:

❗ Hallucinations
Sometimes the AI suggested odd meal combinations or unrealistic workouts (like 500 pushups in a session). While few-shot prompting reduced this, hallucinations still pop up.

📚 Lack of Scientific Rigor
The AI doesn’t understand human physiology like a certified trainer or dietitian. There’s no deep validation of macronutrient splits or progressive overload principles unless explicitly prompted.

💬 Generalization
Without personal health data (e.g., heart rate, metabolism), plans are generic. The prompts can’t fully replace human expertise—yet.

🚀 What’s Next?
I’m excited to push this idea even further. Here’s what I’m thinking for future versions:

⌚ Wearable Integration
Pull data from smartwatches (Fitbit, Garmin, Apple Watch) to adapt workouts based on real-time biometrics like heart rate, recovery, and sleep.

😊 Emotion & Mood Detection
Use sentiment analysis from chat or voice to recommend workouts that match your energy—e.g., yoga for stress or HIIT for low mood days.

📊 Progress Tracking & Feedback Loops
Incorporate user feedback to fine-tune plans dynamically, enabling a closed-loop learning system using reinforcement techniques.

🎯 Final Thoughts
This capstone project was an incredible way to blend health, AI, and user-centric design. Building FitFusion helped me dive deep into GenAI’s practical applications while thinking critically about its limitations and potential.

I hope it inspires others to use AI not just for text or code—but for real impact in people’s lives.

Thanks to the Google GenAI Intensive for the tools, inspiration, and the deadline to make it all happen 😄

✍️ Built with Python, Kaggle, and Google’s GenAI tools.
👀 View the code notebook here 🔗
📹 YouTube demo coming soon!

Smart Content Planner Application

Harshada Javeri — Sun, 23 Mar 2025 18:30:13 +0000

This is a submission for the KendoReact Free Components Challenge.

What I Built

The Smart Content Planner application is designed to empower content creators by providing a streamlined platform for planning, organizing, and managing their content. Built with a modern frontend using Vite and JavaScript, and a robust backend powered by Python and Django REST, this application offers a seamless user experience. Additionally, the deployment on AWS ensures scalability and reliability. This document outlines the architecture, features, and deployment strategies of the Smart Content Planner application.

Demo

https://github.com/analyst-harshada/smart-content-app

Features

✅ Post Scheduling – Schedule social media content with a visual planner.
✅ Content Calendar – Drag & drop posts across dates and platforms.
✅ Smart Asset Organizer – Manage and categorize media assets efficiently.
✅ AI-Powered Caption Suggestions – Generate engaging captions effortlessly.
✅ Rich Text Editor – Format posts before publishing.
✅ Analytics Dashboard – Track post engagement & insights.

KendoReact Experience

I leveraged KendoReact Free Components to enhance the app's usability and efficiency:

Scheduler Component → Enables seamless post scheduling.

Grid Component → Displays and manages scheduled posts.

Editor Component → Enhances post formatting with rich-text support.

Charts Component → Visualizes engagement analytics.

Drag & Drop → Makes content reorganization intuitive.

The integration of KendoReact components significantly reduced development time, ensuring a professional and modern UI.

AIm to Impress

I integrated Generative AI to enhance content creation:

Smart Caption Generator – Uses AI to suggest compelling captions based on trends.

Post Performance Prediction – Analyzes and predicts engagement based on historical data.

Delightfully Designed

I utilized:

Kendo UI Figma Kits → To prototype and maintain design consistency.

Progress ThemeBuilder → To customize UI components to match branding.

DEV Community: Harshada Javeri