Managed model serving platform: Worked on resource attribution and auto-scaling of model serving services. Saved >$1.5M opex cost.

GenAI for search: Worked on LinkedIn's embedding-based retrieval (EBR) within a centralized search service.

Presto router (https://github.com/prestodb/presto/tree/master/presto-router)

Presto-Iceberg connector (https://github.com/prestodb/presto/tree/master/presto-iceberg)

Presto ML-based query predictor (https://github.com/prestodb/presto-query-predictor)

Led a unified cloud-native data access solution with Alluxio for end-to-end machine learning pipelines, which overcomes I/O challenges and improves GPU utilization to 90%+ in enterprise-grade ML workloads.

Led the collaboration with Uber on the adoption of caching to improve data analytics (Presto) & machine learning training (Ray) performance and cost efficiency.

Led Twitter's Presto federation (10+ Presto clusters with 3000+ nodes, processing ~10 PBs of data daily).

Drove the project of creating an end-to-end machine learning pipeline learning from request logs to forecast resource usages (92%+ accuracy) of SQL queries of Presto and storing data exploration jobs in JupyterLab.

Contributed to the router and scheduler to improve system performance (P99 query queued time decreased by 90%).

Worked on the low-code ML project to advocate using BigQuery ML to simplify ML pipelines on multiple components such as model training and feature store. Helped re-build Twitter's notification ML models and reduce the dislike rate by 2+%.

Led Twitter's Zeppelin notebook service (~400 weekly active users).

Drove the migration of on-premises Twitter's Zeppelin, to Kubernetes in the cloud (GCP) in a move-and-improve strategy.

Worked on a unified ingestion web service for Apache Druid, a real-time analytics database, to manage data ingestion jobs.

Contributed to the company-wide graph analytics project and evaluated a hybrid architectural design with Spark and Neo4j for next-generation large-scale graph analytics.

Implemented an MBaaS (Mobile backend as a service) layer on top of heterogeneous data stores (table store, object store, etc.) to enforce the strong consistency of CURD operations on composite data objects.