AI -> agentic workflows, applied LLM systems, evaluation loops
SWE -> clean architecture, backend systems, product-grade tooling
Automation -> reliable pipelines, developer workflows, operational leverage| Build | small useful systems, then harden what survives contact with reality |
| Think | interfaces first, failure modes early, tests where risk is highest |
| Ship | clear scope, measurable behavior, boring reliability |
| Project | What it does | Stack | CI |
|---|---|---|---|
| tiny-unet | Minimal, readable U-Net for cell segmentation — trains from a single command on synthetic data | Python · PyTorch | |
| contrib-calendar | Zero-dependency SVG contribution heatmap (GitHub-style calendar) for any data | TypeScript | |
| pr-size-label | GitHub Action that auto-labels PRs by diff size — lockfile-aware, idempotent | Python · Docker | |
| nvim-config | Fast, lazy-loaded Neovim config: LSP, treesitter, lazygit, polish | Lua | |
| gh-streak | Contribution-streak analytics from the CLI: current/longest streaks and an ASCII year calendar | Python (stdlib-only) | |
| mot-metrics | Lightweight MOTA / MOTP / IDF1 tracking metrics with Hungarian matching | Python (zero deps) | |
| dotfiles | Dev environment: shell config, aliases, and a one-command installer | bash · zsh | — |
| streak-bot | Keeps the contribution graph warm — a random 5–20 commits every day, fully automated (private) | GitHub Actions · bash | — |
flowchart LR
I[Idea] --> P[Prototype]
P --> E[Evaluate]
E --> H[Harden]
H --> S[Ship]
S --> L[Learn]
L --> I
AI • SWE • systems that earn their keep