I have broad experience across applications, infrastructure, and ML. I’ve built systems at scale (1→100) within large tech companies and also led 0→1 ML infrastructure development in nimble startups.
This blog collects my notes on building AI infrastructure — GPU kernels, distributed training, LLM inference, and ML platforms. Read more about me or browse my projects.
This blog post covers Deepseek V4 model architecture and inference.
Modern asyncio in one sitting: why it is the default for high-throughput Python, the event-loop mental model, the API groups (run → define → schedule → wait → bound → coordinate → escape), when to use which, common concurrency patterns, and two worked programs — a data loader and a batching server.
This blog post covers the MLOps platform design principles for research lab.
This blog post covers the AI compute scaling out network architectures.
This blog post covers the LLM inference context parallel.
This blog post covers the from NCCL to DTensor: The Anatomy of PyTorch Distributed.
This blog post covers the numerics in deep learning.
This blog post covers the quantization in deep learning.
This blog post covers the math behind deep learning - part 1.
A deep dive into the transformer decoder architecture used in LLMs — covering math equations, FLOP calculations, and weight/activation memory analysis for each layer.