About
I'm a computer science student. Most of what I build is a reimplementation of something that already exists, because rebuilding a system is the only reliable way I've found to learn what its design is actually load-bearing for.
The pattern that keeps repeating: the interesting failure is never in the part I expected. A rate limiter that measured itself as correct while sending a hundred simultaneous requests. A memory manager whose instrumentation became the largest cost in the system. In both cases every component was right and the composition was wrong, which is a hard thing to catch from inside a single process.
So I've been writing them down. That's what the writing is.
Current work
- nodejs-k8s: Kubernetes' core APIs in Node, compatible enough that the real
kubectland standard YAML work against it unmodified. - Kestrel: a macOS browser on WKWebView with a memory-budget tab scheduler. It holds a budget you set by demoting background tabs down a ladder of measured states. Not public yet.
- social-deduction-bench: measuring LLM deception and deception-detection per turn and corrected for chance, rather than by win rate.