Backend & DevOps Engineer. I build CLI tools, databases, automations, and containerized applications using Python, Java, C++, Docker, CI/CD, and MySQL. Three CLI tools I currently maintain, all open source.
standup-bot — turns git history into a daily standup using a local LLM (Ollama) or a free cloud one (Groq).
Python · SQLite, WAL mode · Rich
v0.2.3 had a real bug: a timeout was passed as a model parameter instead of an HTTP timeout, so Ollama silently ignored it and requests could hang forever. Fixed by moving it into ollama.Client(...), with a regression test so it stays fixed. CI gates on ruff, mypy, bandit, and pip-audit before tests run; all SQL is parameterized, and secrets like tokens and API keys are redacted from every log and terminal output.
env-auditor — diffs the env vars your code actually references against .env.example, across six languages, catching what's undocumented, stale, or missing a default.
Python · zero runtime dependencies
Written defensively since it scans code it doesn't control: long lines are skipped to avoid ReDoS, symlinks are never followed, and --exclude paths can't escape the scan root. 141 tests, 85% coverage enforced in CI, tested across three OSes and three Python versions, with mypy --strict as a hard gate.
chess-review-bot — reviews a PR's diff and labels each file with chess Game Review vocabulary (Brilliant, Blunder...), using a deterministic rule table instead of an LLM call, reasoning documented in an ADR.
Python · GitHub Action + PyPI CLI
Shipped broken: the Action doubled its own path and failed every run with exit code 2 before classifying anything, caught immediately by dogfooding it on this repo's own PRs and fixed in v0.1.1. Published to PyPI via OIDC; secrets are stripped from every subprocess call before it touches git.
Also built position-evaluator for a weekend hackathon: a MySQL-backed Flask app using Gemini structured output, where the lessons on getting reliable output out of an LLM led directly to chess-review-bot's classifier above being deterministic instead.
Open source contributions — merged three fixes into Termstory on three consecutive days, each in a different subsystem. The circuit breaker fix alone went through five rounds of review, catching a race condition and a silent mutation trap in a backward-compat shim along the way. Two more fixes followed over the next two weeks.
Root-caused four separate bugs behind AynOps' header analyzer disagreeing with browser DevTools.
Hardened composable-data-stack's Docker image; pinned to a digest, moved off root; validated by actually building and running it, and separately closed a .env leak into the build context.
Fixed a session-hijack path in thumper and added its ruff linting. Three merged fixes to odys, an energy-optimization model, including one where the linked issue's own suggested fix was mathematically wrong. More merged PRs: full list.
Stats
X / Twitter — build-in-public updates.