Oct 2014 – Jun 2025 · 10 yrs 8 mos · Menlo Park

• In recent years, I led to redesign search tech stack with LLM including its QU and relevance stack. We developed a set of customized and highly efficient AI/LLM across stack, and successfully deployed them to our production with billions of queries per day. I also led Meta AI search integration and transited into AI search product.
• I started leading search relevance org from 2019 and the whole search org from 2021 including search platform, NLP, search relevance, search product, nullstate, typeahead and mention etc.
• To 2019, as senior staff and TLM2, led search retrieval and ranking
• To 2016, as staff engineer, worked on ads delivery, ads outcome prediction and bid suggestions, real time ads targeting, real time content insight platform etc.

Jan 2013 – Jan 2014 · 1 yr · San Francisco Bay Area

• Within one year, drove a team with a few awesome peers to build a large-scale deep learning platform from the scratch for Samsung's DL-based speech recognition, image processing and NLP . It was build based on several hundreds of machines and on top oh Hadoop Yarn framework. We implemented both model-parallel and data-parallel training with distributed parameter servers. With the above platform, we implemented CNN, fully connected neural network.

Jan 2011 – Jan 2013 · 2 yrs · Hillsboro, Oregon

• Big data frameworks:
• Designed and implemented a machine learning based tuner to automatically tune Apache Hadoop framework (Intel branded Smart Tuner)
• Optimized Hive to speed up MapReduce based interactive processing
• Optimized Hadoop MapReduce stack to reduce job launch time
• Worked on a large scale graph construction library called graphbuilder (Intel branded GraphBuilder) for GraphLab
• Worked on large scale graph databases (e.g. Titan, Neo4j)
• Worked on large scale graph processing engines (GraphLab, Giraph)
• Graph Analytic live demo at Intel CTO's keynote in IDF 2013
• System architecture work in data center, network and storage I/Os:
• worked on design and implementation of micro-servers in data centers. The new servers pack lots of wimpy cores, disaggregate network and storage devices and interconnect them using high optical fabric LightPeak fabric.
• developed high performance in-kernel block level storage stack for micro-servers. The stack allows wimpy cores to remotely boot OS and efficiently share disaggregated SSD/HDD storages over LightPeak. The stack is much more efficient than existing stacks like ISCSI.
• Worked on NVMe driver
• Worked on RDMA-based NVMe storage stack using InfiniBand
• Worked on GFS over RDMA-based NVMe device for intermediate data of MapReduce
• Several research papers, invited talks, US patents, two times Division Recognition Awards.

Jan 2010 – Jan 2010 · 0 mo · Hillsboro, Oregon

• Worked on I/O designs and implementations for new micro-server in large scale data centers

Jan 2008 – Jan 2008 · 0 mo · Hillsboro, Oregon

• Integrated on-die high speed NIC research: worked on Niagara2 processor and developed some cache features in full system simulators

Jan 2007 – Jan 2007 · 0 mo · Hillsboro, Oregon

• Developed several optimizations in Xen hypervisor and driver domain to improve virtualized I/O processing performance with 10GbE

Jan 2002 – Jan 2006 · 4 yrs

• I spent around four years working on open source Linux development on various Intel platforms like Xeon and Xscale. Key summary of my work are below:
• Linux Graphic driver development
• Reference at http://lists.freedesktop.org/archives/xorg/2006-August/017270.html
• Real-time POSIX IPC development
• Reference at https://lists.linux-foundation.org/pipermail/robustmutexes/2004-August/001561.html
• High Availability Linux Cluster design and developments
• Reference at http://marc.theaimsgroup.com/?l=osdlcluster&m=108933427704662&q=p3
• Linux Kernel Enabling: enabled new features of Xeon servers in kernel and fixed bugs for local OS vendors
• Embedded System: developed a Watchdog device driver for Xscale and co-ported a Linux Kernel Debugger tool (KDB) to Xscale