Backend engineer focused on Go infrastructure and LLM tooling. I work mainly on Prometheus (CNCF), Kubernetes tooling, and Python ML infrastructure — and I run Windows as a daily driver, which is how I keep running into the bugs that never show up on Linux CI.
Stack: Go · Python · Kubernetes · Docker · Prometheus · GitHub Actions
Merged pull requests, with the defect each one removed:
| Project | PR | What it fixed |
|---|---|---|
| juanfont/headscale | #3650 | Pre-0.23 databases without a migrations table were migrated blindly; now they are rejected with a clear error instead of being corrupted. |
| pranshuparmar/witr | #242 | The root main.go was a tracked symlink. Git for Windows checks symlinks out as plain text, so go build failed on a real clone while CI stayed green. Replaced it with a real entry point and added a lint guard for tracked symlinks. |
| dagucloud/dagu | #3554 | mark_success was ignored for non-retryable exit codes, so marked steps still failed the workflow. |
| dagucloud/dagu | #3578 | Regenerated llms.txt and added a CI check that keeps it in sync with the skills tree. |
| k8sgpt-ai/k8sgpt-operator | #855 | Added tolerations to the K8sGPT CRD and Helm chart so the controller can schedule on tainted nodes. |
| bmf-san/ggc | #542 | The config and command test suites relied on Unix-only exec mocks; they now pass on Windows. |
Most of my fixes start from a failure I hit locally: a Windows checkout that will not compile, a GBK locale that mangles a template file, a timestamp that drifts by a few nanoseconds. I reproduce it, write the failing test first, and open the PR only after the affected suites, vet and formatting are all green.
A coding-agent runtime you can audit. Every tool call, model reply and MCP request lands in a hash-chained JSONL log; veritas replay verifies the chain, re-runs the tools, and diffs each fresh result by SHA-256 — a run is verified, not trusted.
- 141 tests, zero runtime dependencies, CI matrix Linux / Windows / macOS × Python 3.10–3.13
- Built for real Windows environments: UTF-8 → GBK → Latin-1 fallback, BOM and CRLF preserved
- Includes a tamper demo that edits a session and shows replay rejecting it before any tool executes
Turns lm-evaluation-harness output into a CI verdict. Given before / after result files, it decides whether a benchmark actually regressed: a relative threshold plus a standard-error band, not "the third decimal went down, turn the build red".
- 107 tests, zero runtime dependencies (
dependencies = []) - Ships an offline demo where MMLU stays inside the noise band and GSM8K correctly fails the gate