FormFit

FormFit is a watch-first rehabilitation and form-feedback platform built around Apple Watch motion capture, an iPhone companion app, and a lightweight machine learning pipeline for movement-quality scoring.

This repository has been updated from the original watch-only collector into a fuller Spring 2026 snapshot of the project: branded phone and watch UI, end-to-end watch-to-phone session transfer, app-side session storage, and the PyTorch-to-Core ML model pipeline used to prepare future in-app scoring.

Highlights

• End-to-end Apple Watch motion collection using CMDeviceMotion
• iPhone companion app for receiving transferred sessions and storing them in-app
• Branded phone and watch UI with synchronized visual language and app icons
• Data pipeline for collecting, labeling, splitting, and training movement data
• 1D CNN training workflow with Core ML export included in the repo
• EOH 2026: FormFit earned 2nd Place in Most Industry Impact after presenting to 30,000+ attendees

System Overview

FormFit currently spans three connected layers:

1. Watch collection The Apple Watch records user acceleration, gyroscope readings, and roll/pitch/yaw orientation data during rehab or strength movements.

2. Phone experience The iPhone app receives session files from the watch, preserves the raw CSV data, and imports sessions into app-owned storage so the data is ready for analysis and future scoring flows.

3. ML pipeline The ml/ section contains the training code, labels, split rep files, trained weights, training curve, and exported Core ML package used to support movement-quality scoring.

Repository Structure

.
├── ml/
│   ├── CNN.py
│   ├── FormFitModel.mlpackage/
│   ├── best_model.pth
│   ├── formfit-data/
│   ├── formfit-labels.csv
│   ├── training_history.png
│   ├── README.md
│   └── requirements.txt
├── tools/
│   └── generate_formfit_icons.swift
└── watchos/
    ├── FormFitWatch/
    │   ├── FormFitPhone/
    │   ├── FormFitWatch Watch App/
    │   └── FormFitWatch.xcodeproj
    └── README.md

What Is Included

Apple Watch + iPhone app

• Watch motion capture flow with start/stop/save controls
• WatchConnectivity transfer from watch to phone
• iPhone-side import and persistence of workout sessions
• UI work for home, workout, progress, profile, tutorial, and leaderboard screens
• Themed watch and phone styling plus branded FormFit app icons
• Foundations for future coaching and scoring workflows

Machine learning workflow

• Rep-window training data under ml/formfit-data/
• Label file under ml/formfit-labels.csv
• PyTorch CNN training and scoring script in ml/CNN.py
• Trained checkpoint in ml/best_model.pth
• Exported Core ML package in ml/FormFitModel.mlpackage
• Saved training curve in ml/training_history.png

Running The App

Open the Xcode project at:

watchos/FormFitWatch/FormFitWatch.xcodeproj

Then:

1. Select the FormFitPhone scheme if you want the paired iPhone app flow.
2. Select the watch app scheme if you want to work directly on watch collection.
3. Run on a real iPhone + Apple Watch pair for sensor collection and transfer testing.

More detailed notes live in watchos/README.md.

Running The ML Pipeline

From the ml/ directory:

python3 -m pip install -r requirements.txt
python3 CNN.py

That script:

• loads split rep CSV files
• trains the 1D CNN
• saves the best weights
• writes a training history plot
• evaluates validation accuracy
• exports the model as a Core ML package

More details live in ml/README.md.

Current Status

This repo reflects a strong prototype milestone:

• collection is working
• watch-to-phone transfer is working
• app-side storage is working
• the UI is far beyond the original blank collector prototype
• the ML training/export pipeline is present

The next major layer is tightening the connection between imported sessions and live in-app scoring / coaching.

Acknowledgment

FormFit was built through months of iteration across sensing, UI, data engineering, and model experimentation. This repository is meant to present that work as a cohesive system instead of only the earliest collector prototype.