Served as tech lead of our Ray-based model-serving system, which we grew into the company-wide platform for model hosting and external LLM access; designed the sharding system that uses rendezvous hashing to spread tens of thousands of models across several physical clusters behind one logical service.
Reduced our organization's cloud spend by 40% while improving latency and reliability across services within and on top of the AI platform, by optimizing how they use infrastructure rather than just cutting the capacity they run on.
Helped design and build an Argo CD-based deployment system for the AI platform and the ~150 services on our Kubernetes clusters, cutting pull-request-to-production time from eight hours to three minutes, with Rego policy checks on rendered Helm and Kustomize manifests, so bad configuration fails in CI.
Senior Machine Learning Engineer
November 2021 – November 2023
Led the original design and build of the model-serving system, 50x cheaper than the one it replaced; won a Workday Innovator Award for it in 2023.
Helped design and build the monorepo that brought our shared inference service framework to dozens of services and hundreds of engineers.
Migrated a service's model-building and ETL pipelines from our colocated data center to AWS, trading a fixed hardware footprint for elastic capacity; won one of 25 quarterly Outstanding Contributor Awards for the work in Fall 2022.
Machine Learning Engineer
November 2020 – November 2021
Factored a shareable inference service framework out of an existing service, giving later services a common base to build on.
Built tooling for defining our alerts and dashboards programmatically, so that observability was reviewed and versioned like the rest of our code.
Worked with researchers to turn their science into automated high-throughput workflows, generalizing one-off solutions into shared tools like Dask-CHTC.
Modernized how HTCondor taught and tested itself: rewrote the Python bindings tutorials for users rather than administrators, and built Ornithology, a pytest-based alternative to the existing Perl-based testing framework.
Project Assistant
Summer 2018 – Fall 2019
Primary developer of HTMap and htcondor.dags (now part of HTCondor itself), Python packages for describing and running HTCondor workflows.
Taught discussion and laboratory sections ranging from electromagnetism and optics labs to physics in the arts, and revised the manuals and supplementary materials for the advanced lab courses; won the department's Joseph R. Dillinger Award for Teaching Excellence in 2018.