Led data infrastructure team as Tech lead under ML platform, architecting lakehouse-based platform services to streamline Snap training data management and expand data accessibility

• Led lakehouse team to launch ad-hoc query services via customized Apache Trino, integrating with Snap security infrastructure, Kafka, Prometheus, Looker, Vertex AI, and additional ecosystem

• Launched in-place registration service, facilitating zero-copy registration of PB-scale parquet files (from BigQuery) in Apache Iceberg table format using Spark, enabling unified analytics

• Improved Trino's Iceberg connector performance, reducing production ad-hoc query runtime by 5X

Led Wish's central Data Platform team, overseeing online and offline storage systems, core data infrastructure, workflow orchestration, and analytical services — with primary focus on driving cost efficiency, reliability, scalability, and adoption of emerging technologies. Achieved 70% infrastructure cost reduction and millions in annual savings by optimization and migration from AWS EMR to self-managed big data ecosystem on K8s.

Compute and Streaming Infrastructure

• Self managed Spark on K8s, customized Presto/Trino on K8s and AWS EMR

• Log ingestion and MongoDB ETL using CDC, Kafka, Flink, Spark

Online Database Infrastructure

• On-prem MongoDB deployment across multiple self-operated data centers

• Self managed Vitess/MySQL on K8s and RocksDB on K8s

Analytical and Orchestration Infrastructure

• Self managed Apache Superset on K8s, Apache Airflow on K8s and Data Catalog/Datahub on K8s

Data Infrastructure

Presto developer, a distributed SQL query engine for big data. Focus on query optimization, memory/CPU/IO efficiency, large batch query processing, data serialization. system reliability and scalability.

• Developed CPU efficient data file format with 45% CPU improvement

• Optimized Jetty thread pool using async-IO and other techniques, achieved 30X thread saving

• Reduced peak memory utilization in TableFinishOperator by 88%

• Optimized UnionAll operator from a single node execution plan to parallel processing

• Optimized Hive table bucketing, reduced empty file creation by 35%

• Reduced 30% user error reach by optimizing geo-spatial queries and intelligent throttling of high stage queries

• Improved ORC file writer by implementing sampling-based optimization, achieved 5X latency reduction

Location Platform Infrastructure

• Designed and launched geo-spatial indexing services leveraging distributed KV storage, reducing query latency by 10 and powering millions of QPS at scale

Research and development engineer at HPE Vertica, a MPP analytical database management system, focusing on query plan optimizations, system scalability and fault tolerance

• Developed node pruning strategy to improve Spark Vertica connector, utilizing segment distribution, hash range expressions and locality-aware parallel connections

• Optimized key value query via single node planning instead of distributed query plan to improve query performance and cluster scalability under high concurrency

• Experience in PostgreSQL internals

Full stack web development