An engineering lab, not a content farm.

What this is

Inference Lab publishes deep, measured engineering content on LLM inference performance: how vLLM actually schedules requests, what the KV cache really costs, how attention kernels spend their cycles, and how to profile and quantize without lying to yourself.

The bar for every post: a senior engineer should be able to rerun the numbers. If a claim can't be benchmarked, it gets labeled as opinion. If it can, the benchmark ships with the post.

Who writes it

Bio coming shortly — the short version: a distributed-systems engineer of twenty-plus years, CTO and co-founder of an AI inference infrastructure startup, and a veteran course author.

What's coming

• Writing — long-form posts on inference internals, starting with the vLLM request lifecycle.
• Courses — the first is KV Cache: The Complete Mental Model, with a flagship performance-engineering course to follow.
• Open source — annotated-kernels: teaching-grade, benchmarked GPU kernel implementations.