Ran around and did lots of things in the last four years here --
• Implemented automatic model partitioning for multi-gpu inference; built this on top of a system created by our R&D team creating a Virtual Tensor abstraction that streamed PyTorch ops to remote workers. (Python)
• Closely interfaced with open-source PyTorch 2.0 developments, filed issues, and contributed fixes. (Python)
• Handled, ran, profiled, and debugged customer models and open-source models directly, in various frameworks including PyTorch, ONNXRuntime, TensorFlow, and TVM.
• Drove model deployment and accuracy validation initiatives, both internally and for external hardware companies. (Python)
• Significantly reduced model optimization strategy search time via design of a tuning records caching service. (Rust, Python)
• Bootstrapped visualizations and frontend interactions, from scratch. (React, JS, vega)
• Helped bootstrap the SaaS platform, from scratch. Defined database schemas, created APIs, defined protobufs, worked on all components of the whole e2e flow: client -> api server -> scheduler -> worker. (Rust)
• Did a lot of random terraform.
• Iterated through alternative prototypes to discover product-market fit.
• Hacked on the TVM compiler -- improved coverage and added metadata passes. (C++)
• Designed and created a weekly integration model testing service. (Python, Kubernetes)