At TikTok, I lead the real-time Kafka and Flink ingestion workstream for For You Page ranking, landing 10B+ daily events in Delta Lake with sub-second freshness. This work contributed to a +0.6% watch-time lift in online A/B tests.
I designed Databricks lakehouse workflows that replaced full recomputes with incremental merges, bringing Silver and Gold table refreshes down from hours to minutes. I also developed multimodal embedding pipelines for 500M+ items and rolled out six model upgrades with zero retrieval downtime.
At TikTok, I delivered an online/offline feature store using Delta Lake, Redis, and Cassandra, reducing training/serving feature drift by 40%. I also optimized Flink and Kafka infrastructure, cutting costs by 25% and eliminating recurring checkpoint-timeout pages during peak traffic.
At DataLab USA, I architected DLINK and CLINK consumer identity resolution platforms using Apache Spark, Databricks, and SQL Server. CDC-based ingestion unified 25+ sources covering 200M+ consumers, while explainable matching supported FCRA-governed use cases.
