At OpenAI, I built and led core components of the post-training pipeline used to train GPT-4.5, GPT-4.1, and GPT-4o. I designed an RL framework for LLM post-training, including training-loop components and rollout generation.
I also built modeling infrastructure behind Advanced Voice Mode, working across the audio encoder and decoder pipeline for real-time speech. My work supported multimodal development for GPT-4o and GPT-4o-mini, as well as training infrastructure for GPT-4 Turbo.
At Google, I led large-scale model inference and built production serving infrastructure for Vertex AI, powering ML features in Gmail and Google Docs. I also contributed to T5X, PaxML, Orbax, and TensorFlow DTensor, including foundational distributed training capabilities.

