Image
When Her Body Became the Dataset: Building an AI Recovery Coach After Two Hip Surgeries
Gina Mancuso had 25 days between finishing her code and toeing the start line at the New York Hyrox — her return to the global fitness racing competition after two hip surgeries in four months. She is a data science student, a high school statistics teacher, and a single mom. Constraints are nothing new to her. But one she couldn't live with was this: she had more biometric data than ever before, and absolutely no clear picture of what to do with it.
Her Oura Ring had opinions. Her Garmin had different ones. Social media had 17 more, none of them written for a post-surgical athlete trying to train smart without setting herself back. So she did what data scientists do when the tools don't exist yet: she built it herself. The Mancuso Method is a personal AI recovery coach that reads her wearables, remembers what she ate for dinner, knows which hip is further along in rehab, and delivers a personalized morning briefing every day. In the text that follows, Mancuso shares how it works — and what building it revealed about what real data engineering actually looks like in practice.
I am a competitive Hyrox athlete, a high school statistics and economics teacher, a single mom, and a first-year student in UVA’s online M.S. in Data Science program. In December 2025, I had my first hip surgery. In March 2026, I had my second. Somewhere between physical therapy appointments, lesson plans, and grad school, I realized I had a data problem that data science could solve.
The Problem
If you use wearable devices, you probably know this feeling: I wake up, check my Oura readiness score, glance at my Garmin HRV, scroll through three different apps, and somehow end up more confused on my status and what I should do than when I started. Social media adds another layer with recovery protocols, nutrition advice, training philosophies, and coaches marketing themselves. And of course, they all contradict each other, none of it accounting for the fact that I am a competitive athlete who had two hip surgeries within four months and am actively rehabbing in hopes of doing the New York Hyrox in June.
For adult athletes, especially, the gap between data and decision-making is enormous. The data exists. The wearables are tracking everything. But translating a readiness score of 77 and a heart rate variability (HRV) of 41 milliseconds into a concrete answer for “what should I do today and what should I eat?” requires either an expensive coach or a willingness to guess.
This is when I decided to stop guessing and build something instead.
Project Overview
Mancuso Method (MM) is a personal AI recovery coach that pulls live biometric data from the Garmin Connect and Oura Ring APIs, synthesizes it with daily nutrition logs, and delivers a personalized morning briefing powered by Anthropic’s Claude AI. The system knows my surgical history, my training schedule, my PT protocol, and even what I ate for dinner last night, because it read my conversation from the day before.
The project is built entirely in Python, runs locally on my MacBook, and has a browser-based chat interface I can use instead of a terminal. It is the most personal application of data science I have ever built.
Data Sources
- Garmin Connect API — activity data (indoor cycling, strength, elliptical, rowing), sleep metrics, HRV, body battery, and daily nutrition logs entered through the Garmin app.
- Oura Ring API — readiness score, temperature deviation, and sleep stage data providing a second source of recovery signal independent of Garmin.
- Conversation memory — daily coaching conversations are saved as JSON files and loaded the following morning, giving the AI coach continuity across sessions. This implements a lightweight RAG (retrieval-augmented generation) pattern without a vector database.
Skills Demonstrated
API Integration & Data Pipeline Architecture: Built authenticated connections to two wearable device APIs simultaneously, handling token caching, rate limiting, and pagination across date ranges. This was the direct application of skills developed in UVA’s Data Pipeline Architecture course.
Data Wrangling & Merging: Each data source returns different date formats and time zones. Merging Garmin sleep data, Oura readiness scores, and activity logs into a single coherent daily record required careful datetime normalization and join logic using pandas — a problem that took longer to solve than anticipated and produced real lessons about type consistency in merges.
LLM Prompt Engineering: The AI coaching quality depends entirely on how data is structured in the prompt. I built a detailed athlete profile, injected live biometric signals, and included yesterday’s conversation as context — enabling the model to reference specific foods eaten, sessions completed, and PT feedback from the prior day.
System Design: The final architecture separates concerns cleanly: a Flask API serves live data, a Python briefing engine handles coaching logic, and a standalone HTML interface connects both. This makes each component independently testable and replaceable.
Tools:
- Python 3.12 (conda environment)
- pandas, requests, python-dotenv, Flask, flask-cors
- Garmin Connect Python SDK (garminconnect)
- Oura Ring REST API v2
- Anthropic Claude API (claude-sonnet)
- VS Code + Windsurf
Image
What It Actually Does
Every morning I run the briefing and receive a personalized coaching summary that includes my recovery status, what the biometric data says about training readiness, a specific training recommendation for the day, and nutrition guidance based on yesterday’s actual intake. Then I can chat with the coach.
The coach knows I had a 396-calorie deficit two days ago. It knows my left hip is seven weeks post-op and my right hip is five months post-op. It knows I had PT on Tuesday morning and that my hip was tight after the rower and added in a stretch routine for the day. It knows I was planning to have chicken, sweet potato, and broccoli for dinner and that I have to have tortilla chips every day, because they are my favorite.
That level of context does not come from a generic fitness app. It comes from data engineering.
Results and Reflection
The system has been running daily for over two weeks. The coaching quality is genuinely high, specific, data-grounded, and contextually aware in a way that generic AI assistants cannot replicate without the pipeline behind it.
The project also revealed something important about real-world data work: the hardest problems were not the AI integration. They were the messy, unglamorous data issues like date type mismatches breaking joins, Garmin rate limits interrupting data fetches, API tokens expiring at inconvenient times. Those are the problems data engineers actually solve every day.
With 25 days to my debut back in Hyrox with two healing hips, I am racing. And my AI coach, built from what I learned in a UVA data pipeline course, is with me to achieve my goal.
Check out the project on GitHub.
Learn more about the part-time, 100% online M.S. in Data Science at the University of Virginia. Request more information, connect with Admissions, or start your application today.
