I’m a Tech Lead Manager (TLM) on the PyTorch TPU team at Google, focused on optimizing and scaling PyTorch models for training on TPUs.

Previously I was TLM on the Vertex Prediction Performance team, where we focused on optimizing GPU and TPU models for inference.

I’m especially interested in learning how to apply new agentic capabilities to improve both the development process and model optimization.

This blog is where I write up what I learn along the way: ML performance and profiling, model training and inference, accelerators and compilers, and the developer tooling that ties it all together.

(You’ll see my name two ways: Oleksiy is the official transliteration from Ukrainian; Aleksey is the spelling I’ve used most over the years.)

A few things I’ve built

• tpu_performance_autoresearch_wiki — a wiki built with the Karpathy method, holding TPU performance-optimization knowledge and hooked up to an autoresearch optimization engine.
• xprof-mcp — an MCP server for the open-source XProf profiler.
• jupyter-ai-agent-magics — flexible AI agent capabilities for Jupyter.
• tensorflow_serving_benchmark — a tool for benchmarking TensorFlow Serving.
• gcp_runner — running nbdev projects on Google Cloud Platform.
• human.ai — an experimental human-like virtual assistant.

Elsewhere

• Résumé: Oleksiy Vlasenko
• GitHub: vlasenkoalexey
• LinkedIn: Oleksiy Vlasenko