Lumosity.

We rebuilt Lumosity's personalization engine using a multi-armed bandit approach combined with deep learning models trained on billions of historical gameplay data points. The new system creates a cognitive profile for each user across five domains (memory, attention, flexibility, speed, problem-solving), updates it in real-time after every game, and uses it to construct training sessions that optimally challenge each domain based on the user's current state, recent progress trajectory, and session context (morning vs. evening, time available, recent cognitive load). We also redesigned the progress tracking experience to give users meaningful insight into their cognitive development over time. Instead of abstract scores, we built a visualization system that shows concrete improvements — "Your working memory span has increased from 5 to 7 items over the past month" — grounded in the psychometric data behind each game. The technical migration involved moving from a monolithic Rails backend to a microservices architecture on AWS, enabling real-time model inference at the edge for sub-100ms personalization decisions. We implemented A/B testing infrastructure that lets the research team run controlled experiments on new game mechanics and training algorithms without engineering support.