Palo Alto, California, United States
Search AI - Building auto-embedding pipeline for vector index generation with flexible user-defined model configurations
Branded mission ads: - Led a cross-functional squad (5+ engineers) to built the Brand-Ads Content Understanding Platform (MLLM/LLM/ASR/OCR), processing millions of videos/day to expand ad-targetable inventory and improve match precision, resulting in multi-hundred-million-dollar annual revenue growth. - Built high-throughput offline and online evaluation pipelines (Spark + Flink + storage) providing near real-time (<30 min) model insights to guide launch and training. - Owned the A/B testing and long-running experiment platform for video-ranking models, including configuration, mirrored logging, guardrails, and real-time monitoring for pre-rollout performance evaluation. - Partnered directly with ML engineers and data scientists to define evaluation methodologies, metrics (quality, safety, relevance), and experiment designs: serving as the primary tooling interface between researchers and production systems.
Search ads format optimization: - Owned the full stack (Java/C++ backend, JavaScript UI) for google search ad image formats used by hundreds of millions of users; frequently ran experiments to validate new format ideas. - Designed and built a full-stack research tooling suite for the Ads Formats team: tracking long-running experiments across ad format variants, with dashboards monitoring model prediction metrics (pCTR, click quality) and surfacing regressions and bias over weeks-long windows. Became the primary research tooling suite used by ML engineers and data scientists to evaluate format changes before rollout.
Built and maintained high-performance backend components for core Microsoft Office products, serving hundreds of millions of users globally.
Process business new account opening