Kelvin Zou

Also known as Xuan Kelvin Zou, Xuan Zou

Kelvin Zou is a machine-learning systems engineer whose papers at Apple covered the foundation language models behind Apple Intelligence and the software that trains and serves them. Trained in networking, he helped lead the VeriFlow network-verification project as an undergraduate at the University of Illinois and earned a master's in computer science at Princeton under Jennifer Rexford in 2015. After engineering jobs at Microsoft Azure, Google and ByteDance, where he coauthored the Monolith recommendation system in 2022, he joined Apple as a principal engineer in October 2023. His Apple work includes the AXLearn training system and Parallel Track Transformers, an architecture that cuts GPU synchronization during inference. He is now a research engineer at Meta in Kirkland, Washington.

Now at: Meta
Badges: Bridge
Wayfarer
Links: Google Scholar

Timeline

2012–2013

VeriFlow at Illinois

He studies computer science as an undergraduate at the University of Illinois at Urbana-Champaign, co-advised by P. Brighten Godfrey and Matthew Caesar, and works as a research assistant at the Information Trust Institute from May 2012 to May 2013. As one of the leads of VeriFlow, a tool that checks a network's data plane in real time, he coauthors its NSDI paper in 2013.
2013–2015

Networking research at Princeton

In September 2013, he joins the networking and systems group at Princeton as a research assistant, advised by Jennifer Rexford. With Ryan Beckett, Rexford and David Walker, he coauthors a 2014 assertion language for debugging software-defined networking applications. Princeton awards him a master's degree in computer science in June 2015.
2015–2016

Azure infrastructure at Microsoft

From September 2015 to October 2016, he is a research software engineer on Microsoft Azure infrastructure in Redmond. RAIL, a 2017 NSDI paper that lists him with Microsoft Research, stretches data center optical links past their specified reach and cuts network cost by up to 44% for 40Gbps networks.
2016–2018

Software developer at Google

He works as a software developer at Google in Mountain View from October 2016 to May 2018. Listed with Google, he coauthors Dysco at SIGCOMM 2017, a protocol that steers TCP sessions through chains of middleboxes and can reconfigure a chain while a session runs.
2018–2023

Senior staff engineer at ByteDance

From May 2018 to October 2023, he is a senior staff engineer at ByteDance in Seattle. In 2022, he coauthors Monolith, a real-time recommendation system with a collisionless embedding table, and MeshInsight, a tool that measures the overhead of service meshes.
2023–2025

Foundation models at Apple

Apple hires him in Seattle as a principal engineer in October 2023. He coauthors the 2024 report on the foundation language models behind Apple Intelligence, which include a model of about 3 billion parameters that runs on devices. In July 2025, he coauthors AXLearn, Apple's production system for training large deep learning models, and the 2025 tech report on a server model built on a Parallel-Track Mixture-of-Experts transformer.
February 2026

Parallel Track Transformers

In February 2026, Apple publishes Parallel Track Transformers, a paper he coauthors from work done while at Apple. The architecture targets fast GPU inference, cutting synchronization operations by up to 16 times relative to standard tensor parallelism, and is tested in the TensorRT-LLM and vLLM serving stacks.
As of October 2026

Research engineer at Meta

He is a research engineer at Meta, based in Kirkland, Washington. Machine learning systems, distributed systems and large language models are his listed research areas.