About

Not a CS kid. A statistics major who happens to ship real software.

Stats rigor (multivariate, Bayesian, time series) · Engineering with AI in the loop, and honest about which parts it wrote · Systems work (Linux hardening, self-hosted services, uptime monitoring). The interesting part is where the three overlap.

Focus

Statistics · Data Science

Stage

Undergraduate

Building

Stats tooling · self-hosted infra

Handle

starsdaisuki

Competitions

BirdCLEF+ 2026 — Silver Medal 52 / 4094 top 1.3% solo

Research-grade soundscape classification with a $50,000 prize pool. The public leaderboard was a trap: I picked my final pair for diversity rather than for the best public score, and when the private board reshuffled I went from 88th to 52nd instead of falling. Disciplined cross-validation beat leaderboard-chasing.
NeuroGolf 2026 — Bronze Medal 208 / 2963 top 7.0% solo

ARC abstract-reasoning puzzles, with a twist: you may not submit code. Every task needs a hand-built neural network shipped as an ONNX file, and solving it is only the entry fee — the score is 25 − ln(cost), so the real competition is who can make the network smaller.
MCM/ICM 2026 — Honorable Mention solo

Four days from a blank problem statement to a 27-page paper in English, typeset in LaTeX — problem selection, modelling, code and writing all mine. Six original analytical frameworks built for the submission.

Tech stack

Statistics

RPython (stats stack)Bayesian inferenceTime series (Cryer & Chan)Multivariate analysisEconometrics

Languages / Runtime

GoPython (uv)TypeScriptRust (light)Kotlin (Android, light)Shell

Infra & Networking

Linux hardeningNginxCloudflare Workers / R2WireGuardfail2ban / nftablessystemdDocker

Web & App

AstroTauriVueGradioDeepSeek APIPagefind

What I care about

Minimalism & order. Declarative package management (Homebrew / Scoop / pnpm / uv), ruthless file system hygiene. Recent: 114 → 54 VS Code extensions, full Anaconda residue cleanup.
Learning systems. Daily English log at 1200+ lines, Zotero tagging on every paper, every MCM iteration version-controlled. Progress without instrumentation is invisible.
Cars. 7,000+ km behind the wheel across Tesla / Toyota / Mercedes GLC / Maybach.
Rubik's cube. CFOP, ~12s on 3×3, plus 2-5 / Megaminx / Skewb / Pyraminx / SQ1.

What I don't do

Ship unfinished work. Every project I call "deployed" is actually running and reachable.
Code dumps without comprehension. If I can't explain why it works, I won't claim I built it.
Chase shiny tools for status. I rotate when they stop serving me (iTerm2 → Ghostty, Anaconda → uv), not to add more.