I build AI agents, then I try to break them.

Mathworld is a shared town where kids K-12 play minigames in which the math is the action, and design their own in the Game Lab. The child makes every design decision: the story, the rules, the pictures, the shape of the whole game. A model turns that storyboard into code, and a Socratic Professor bot coaches with questions instead of answers.

Traditional AI game-generation engines rely only on system prompts from end users and automate the rest of the game-development process, producing subpar games and frustrated users. My system puts students in the driver's seat. It gives them more creativity over the parts AI does poorly, and it teaches children through constructionism, using productive struggle.

As part of my AI-tutoring research at UW, I studied AI alignment and goal drift across long chats.

A Socratic algebra tutor is told to guide, never to give answers. In my preliminary testing, forceful overrides were refused immediately, while gradual pressure, where every step looked reasonable, sometimes walked the tutor out of its role without the usual objections.

My hypothesis is the Boiling Frog Effect: alignment training teaches models the tone of an attack, but nothing teaches them its trajectory. Each message passes inspection but the drift is never judged across the whole conversation. If this hypothesis is correct, robustness numbers measured against overt attacks overstate real-world safety, and the fix is a monitor that scores cumulative drift.