HackMIT Contestant Reverse-Engineered a Puzzle's Hidden Model Using 10,000 API Calls

A participant at HackMIT tackled a face-customization puzzle called Face Value, which required finding an avatar configuration that met four simultaneous conditions, including a confidence score of at least 99.9%. Initial manual testing of sliders revealed some patterns but stalled around 60–77% confidence, as the puzzle's scoring model relied on complex multi-feature interactions. The breakthrough came when the contestant used Chrome DevTools to extract the underlying JSON API endpoint, enabling automated querying at hundreds of iterations per minute instead of one per minute by hand. They then applied simulated annealing, coordinate descent, and random restarts to systematically search the configuration space across roughly 10,000 API calls. The project illustrates how reverse-engineering a hidden interface and applying data science techniques can transform an apparently simple UI puzzle into a tractable optimization problem.