Work Book
(Some thoughts on Engineering and ML)

Building a Triathlon Photo Classifier: The Dataset Challenge

Leading up to my first Olympic triathlon, I built a computer vision model to classify race photos into swim, bike, run, and transition categories. Starting with hastily scraped images, my initial ResNet-18 model achieved only 61.6% accuracy. Through systematic data cleaning - removing duplicates, fixing mislabeled photos, and balancing class distributions - I dramatically improved performance to 89.4% accuracy, proving that data quality often matters more than architectural complexity in machine learning projects.