About

I got into programming through machine learning, then got pulled into the systems underneath it: the solvers, the benchmarks, and the tools that tell you how things actually work (or why they’re breaking).

I find something I only understand as a black box, rebuild the smallest version I can measure, then spend a while figuring out why it disagrees with what I expected. That’s usually the part I enjoy most.

Most of it falls into four lanes.

Discovery

Glassbox: recovering equations from raw data.

Simulation

FMMS, a fast-multipole N-body solver in Rust; CelestialSim, RK4 in the browser.

Measurement

freeze-bench, an execution-graded benchmark for coding models.

Tooling

Hark, WTF, and OpenLMlib, things I wanted for my own workflow.

Rust for anything that has to stay running or stay fast: FMMS, Hark. C++ where Glassbox needs the inner loop. Python for the parts that change every day. Go for CLIs.

F1 and most motorsports, fighter jets, rocket engines, and CFD. In all of them the appeal is the same: something moves fast enough that you have to simulate it to see what happens. I daily-drive Arch, which is mostly an excuse to keep rebuilding my desktop.