Production AI · agents · automations · the 2am debugging nobody posts about.
Merged PRs into other people's production codebases: observability, AI gateways, eval tooling, security scanners. This whole block updates itself daily, because hardcoding your own PRs is embarrassing.
In review: instructor · garak · awesome-mcp-security · clearml · grpc · mcp-injection-experiments · langfuse-python · pytorch · sktime · ragas
Tag a repo flagship on GitHub and it shows up here on the next daily run. No hand-editing this file.
| Project | What it does |
|---|---|
| diffprompt | git diff for prompt engineers |
| brok | An MCP that reviews your architecture, estimates real capacity from cited numbers, surfaces the trade-offs you are making, and roasts bad design in a dry voice. Deterministic engine, no model, runs free and local. |
| cacheguard | Drop-in safety layer that stops semantic caches from silently returning wrong answers — with the first independent open benchmark of semantic-cache false-hit rates. Free, local, LiteLLM + OpenTelemetry. |
| OptiQuant | AI-powered quantitative trading platform — ensemble ML (LightGBM + CatBoost + RF), backtesting engine, and live signal scoring. Deployed on AWS EC2 fornerly, now deployed on streamlit cloud |
Working notes that publish themselves straight from a repo: Field Notes (push a .txt, it shows up typeset). Built-for-the-plot side quests: ragbait, iceman, ctxbridge.
Peer reviewed: funny · delusional · ambitious. Fueled by tea, not coffee.