Li Miao

I lead end-to-end multimodal safety and content understanding systems at YouTube. My work spans LLM-as-a-judge labeling, training and evaluating classifiers built on Gemini encoders, and deploying them at scale.

I also deploy Gemini Embeddings across multiple products, including comment search and video-to-image candidate retrieval. I build agentic evaluation workflows to identify embedding failure modes and turn those findings into data feedback loops that inform next-generation Gemini Embedding training.