Three systems I built and measured. Each one has tests and CI, so any number below can be re-run.
- agent-platform — LLM gateway in Go: per-role RBAC, token quota, multi-layer cache, audit trail, fail-closed on saturation. 88 tests; 200 concurrent requests → p50 42.8 ms.
topic:llm-gateway - oom-evidence-lab — 6 kinds of Java OOM reproduced in containers; each death leaves one evidence bundle and a 9-field alert. CI re-runs all 6 cases in 25 s and fails if the docs stop matching.
- ts-nakama-game-demo — Nakama TypeScript runtime in strict mode: controllers, repositories, mappers, typed models, tests and CI.
- pgvector-recall-lab — recall measured against a fixed question set, not eyeballed.
How I work: the model writes proposals; rules, thresholds and the classification path stay in code; a person approves; CI fails the build when the boundary is crossed.
