Atlas Serving Stack
Tech lead — 2023
An inference runtime that cut p99 latency 3× for a 70B model. Open-sourced; now powers serving at three other labs.
I build large-scale ML systems that reach hundreds of millions of people. Over the past decade I've taken models from research notebooks to production infrastructure — training, serving, and the unglamorous reliability work in between. I care about systems simple enough to reason about and fast enough to disappear.
Selected Work
Featherweight Quantization
Author — 2022
A post-training quantization method that holds accuracy at 4-bit. Paper + library, 4k★ on GitHub.
Looking Glass
Creator — 2021
An interactive tool for visualizing attention inside transformer layers.
Experience
Staff ML Engineer
2021–PresentNorthwind AI
Lead the training infrastructure for a 70B-parameter foundation model. Cut training cost 40% with a custom data pipeline and brought time-to-first-token down 3× in production serving.
Senior Software Engineer
2017–2021Meridian
Built the recommendation system powering the home feed for 200M+ daily users, lifting engagement 18%. Owned the online feature store end to end.
Machine Learning Engineer
2014–2017DeepVision
Shipped the company's first on-device vision model for real-time object detection on mobile, running at 30fps under 50MB.
Education
M.S., Computer Science
Stanford University · Focus: Machine Learning · 2012–2014
B.S., Electrical Engineering & CS
UC Berkeley · 2008–2012
Capabilities
Python · PyTorch · JAX · CUDA · Distributed Training · Triton · C++ · Go · Kubernetes · TensorRT · MLOps · Transformers
Recognition
- NeurIPS Best PaperNeurIPS, 2019
- Patent: efficient post-training quantization2020
- Internal Engineering Excellence AwardMeridian, 2018
Beyond work
Bouldering, Generative art, Espresso