Co-developed a three-stage human–AI pipeline to define, label, and deploy prosocial behavior classifiers where no established ground truth existed. Introduced a human-LLM disagreement loop to refine definitions and improve label quality, then deployed a two-tier inference system that cut inference costs by ~70% while maintaining high precision (~0.90). Demonstrated how deployment-aware ML design can unlock new responsible AI system capabilities at platform scale. Authors: Rafal Kocielnik, Min Kim, Penphob (Andrea) Boonyarungsrit, Fereshteh Soltani, Deshawn Sambrano, Animashree Anandkumar, R. Michael Alvarez