fitness-coach-assistant

Inspiration

CrossFit is highly individual, but most athletes still follow generic workouts and receive little feedback between sessions.

What it does

This is an AI teammate for CrossFit athletes. It combines the expertise of a coach, movement specialist, nutritionist, and training analyst to personalize every session based on readiness, injuries, workout history, and real-world conditions. It does not just prescribe a workout, it learns from every session and helps athletes train smarter, recover better, and stay motivated.

How we built it

I used Antigravity IDE to scaffold a multi-agent application using ADK by planning the different steps:

• Historical Workout Ledger & Trend Intelligence: Persistently logs every workout session (date, movements, sets, reps, loads, time, RPE, notes, weather). Enables instant natural language recall (e.g., "What did I do last Tuesday?", "Show my workout on August 15th") and computes longitudinal trends (tonnage volume, RPE progression, movement balance).
• Web UI on Cloud Run via --with_ui: Deploys with the exact same interactive web chat interface as adk web on local, providing an immediate shareable web app on GCP for testing and hackathons.
• 100% Gemini Flash Tier Across All Agents: Standardized on Gemini Flash for low latency, high throughput, and Google AI Studio free tier compatibility.
• Conversational Readiness & Post-Session Check-ins: Pre- and post-workout subjective feedback loops (RPE, soreness hotspots, sleep feel, energy level) to dynamically adapt training.
• Post-Workout Recovery Recipes & Meal Engine: Dietitian Agent recommends practical, delicious post-workout recipes (e.g., 5-min recovery shakes, 20-min refuel bowls) tailored to dietary preferences (Paleo, Vegan, Omnivore, etc.) and the exact workout volume performed.
• User Geolocation Weather Adaptation: Real-time localized weather retrieval based on the athlete's coordinates/city to modify outdoor training (e.g., running vs rowing, hydration adjustments, extreme heat/cold protocols).
• Curated Movement & Technique Knowledge: Official CrossFit and athletic standards, scaling ladders, and verified movement demo videos.
• Gamification & Behavioral Consistency: PR tracking, skill progression trees, and anti-burnout streak systems.
• Security & Privacy First: Envelope encryption using Google Cloud KMS and in-memory PII sanitization.

Challenges we ran into

The gcloud deployment was done following the documentation since I ran out of Antigravity free credits but I found that the documentation was very helpful even if it took me some time to get used to Google cloud deployment and IAM configurations. What was helpful also was the clear debugging messages I got from the IAM issues. Also, since I relied on Antigravity which created a bit of a custom gcp deployment, I had to pass GEMINI_API_KEY as a var env in the deploy script.

Accomplishments that we're proud of

The deployment of a MVP demo which I used to test the different sub-agents, the orchestrator routing properly the requests. Also, I like how Antigravity created automated tests and verification harness which are very helpful in this AI era to make sure we cover properly the behaviour of the application. Furthermore, Antigravity provided the minimum working code in less than half a day.

What we learned

ADK and GCP deployment.

What's next for fitness-coach-assistant

• Proper auth workflow and production deployment.
• Better memory management of user profile and habits.
• If possible, explore how to use MCP Apps for better interactivity with contents/videos of exercise movements.

Built With

• adk
• python