Co-developed the first automated system to detect and classify prosocial behavior in in-game text chat at scale. Led product direction for a novel self-anchored attention model that improved classification accuracy by ~8% in low-data environments. Enabled platforms to move beyond disruptive behavior detection toward identifying, measuring, and reinforcing positive user interactions. Applied in production-adjacent research on Call of Duty with millions of DAUs. Authors: Zhuofang Li, Rafal Kocielnik, Fereshteh Soltani, Penphob (Andrea) Boonyarungsrit, Animashree Anandkumar, R. Michael Alvarez