The shelf · stop 03

Things Iwrote down

Long, illustrated notes about how something actually works — written while I was working it out. Each one is built as its own page, in its own design system, because a piece about GPU kernels should not look like a piece about anything else.

peel one open

every sticker below opens a different-looking page ↓

Serving, from scratch: a big serif title on cream paper, and Flo — a chip with a smiling face — inside a dashed worker-process box with three little envelope requests queued beside it

new

1 September 2026·~40 min

Serving, from scratch

How a model becomes a service, one component at a time — one model, batching, streaming, many models — then a drawboard where you wire your own server and run it.

its own design system — paper & ink · serif, doodles, a GPU mascot called Flo

LLM servingbatchingstreamingread it

29 August 2026·~35 min

The GEMM Scrapbook

One matrix multiply, seven CUDA kernels, and a 70× speedup — documented like a road trip, with the wrong turns left in.

its own design system — engineering notebook · grid paper, red pen, highlighter

CUDAGPUperformanceread it

The next one

it'll look nothing like the one beside it — that's the rule

The only rule

Every piece is its own website — its own type, its own colour, its own physics. Nothing here is a template.

Hand-built HTML, CSS and JS, shipped as one file: the diagrams are drawn on canvas, the charts run on the real numbers, and the wrong turns stay in — that half is usually the useful one.

tell me what to write next

say hi