who i am and how i work
A small lab for shared inference.
The manifesto and the working principles of a two-person inference research lab.
See the proof page
Coconut Labs is a small inference research lab. We work on schedulers, systems notes, and the small pieces of infrastructure that decide whether shared compute behaves like a tool or a tax.
The lab is intentionally small: two engineers, close to the measurements. A result should have a harness. A claim should have a number. A page should leave enough quiet around the thing it is trying to say.
The first public thread is KVWarden: tenant fairness on shared inference. mlxd follows the same line closer to local Apple Silicon inference. Both projects are about making the shared layer more honest under pressure.
How I work
Honest scale
This is one engineer with a regular collaborator. Jay Patel has built alongside me on several of these projects and is credited where he did. No fake-team plurals beyond that.
Specific numbers
A result earns attention when the units are visible. Latency, pressure, and failure modes stay in frame.
Boring infrastructure
Schedulers, harnesses, and traces are not decoration. They are the artifact.
Ship the code first
A post can be elegant. The repository still has to run.
Slow web
Pages should feel like edited pages. Motion is allowed when it carries meaning.
the practice behind the lab
The Library
The Library is the private study shelf the public work is built on, and it holds four things: the waterline atlas, the academy curricula, the essay drafts, and the masterclass source. The rule is one rep a day, fifteen minutes, go one level below what the job requires and write it down well enough to teach.
The shelf lived behind a login on lab hardware, and that serving stack was retired in August 2026, so it is offline rather than gated. What was cleared for the public stays public: the course at masterclass, and the atlas work in the projects gallery. What the lab measured lands in evidence.