Led xFN effort across engineering, data science, research and product design to grow the retention and engagement of Discover (Google's personalized interest feed) that has 400M+ DAUs across 150+ countries. Directly/indirectly manage an org of 20+. Drove projects spanning Content Retrieval, Ranking, Notifications and Features. Built State-of-the-Art Deep Neural Network and Reinforcement Learning based Recommendation Systems to serve 4B+ recommendations/day. Delivered XM+ DAUs through 100+ launches.

+ Built and managed Pinterest’s core advertising serving and data systems. Led the following teams and key initiatives:--- Online Serving: owns the entire ads serving stack, including query expansion, inverted-index based retrieval, online ML scoring, trimming/reranking/blending, auction, and logging. Rebuilt stack from monolithic service to service-oriented architecture to provide 10x scalability. Scaled the system to 40K+ QPS user traffic, 200M+ QPS candidate processing, 100+ concurrent experiments/models, 100+ customized trimmers/rerankers, <250ms p99 latency, and flexible framework that support 50+ ranking/product engineers to continuously develop new features/models/experiments.--- Indexing Systems: aggregates signals from 50+ sources to build inverted/forward indices to power online serving. Rebuilt system from single-box batch solution to distributed batch + incremental + realtime. Scaled system by 500x (to >100M ads documents). Reduced e2e indexing delay for campaign changes by >90% (to <15min).--- Real-Time Budgeting Systems: provides real-time control in online serving to throttle out-of-budget ads and smooth out ads delivery. Rebuilt system in Kafka Streaming and HBase to improve scalability by 100x and e2e delay by >95% (to <10s).--- Data Systems: joins/processes log events and serves aggregated analytic insights data to power advertiser reporting/analytics, data warehousing, and ranking pipelines. Rebuilt pipelines from Hive + MapReduce to Spark + Spark-Streaming based lambda architecture in Scala to process 20TB+ data/day. Improved the e2e delay by >80% for real-time analytics (from >3hrs to <30mins) and daily reports (from >12hrs to <2hrs). Unlocked advertiser spend on the platform by allowing advertisers to make real-time decisioning based on real-time campaign performance stats.+ Build the team from scratch, and grew the team from 3 engineers to 16 engineers in 2 years. Grew and nurtured Tech Leads for each of the 4 teams.

Managed an Engineering team to build Google's first mobile app content crawling + indexing system to enable Google to index and serve in search results all the awesome content hidden in mobile apps. Designed and implemented multiple subsystems including app crawler, app content indexing pipeline, app content understanding engine, and app validation system. Tackled problems spanning across hardware (app simulation), infra (indexing 70B+ mobile pages with O(s) latency) and ML (structured app content extraction). Led team to launch the product for Google I/O in 6/2014.

TL of Google's Social Graph Service which powers the centralized storage and serving of all social graph data at Google spanning across products including Hangout, Groups, Communities, Voice, and G+. Led team to develop the main system on Java and Spanner storage. Scaled the system to 10B+ social connections and 1M+ queries/second with strong consistency and < 12 ms p50 latency across 15 data centers and 70+ clients. Led several major rewrites and migrations.

TL of Google Play Music's Metadata Management Service. Built the system for ingesting and processing the artist/album/track metadata for 30M+ music tracks, managing their relationships, and powering core product features such as search, Scan-and-Match, and recommendations. Built system e2e in C++ and Bigtable storage. Wrote various subsystems including Music Store Service and Music Suggest.

Led an Engineering team to build Search Suggest autocompletion for CJK languages based on their Latin transliterations (Pinyin, Romaji, etc.).

Built various quality eval tools and Search features.