14 years building production systems — 5 of them leading a quant trading team, where correctness and latency left no room for hand-waving. I work AI Native and full-stack now, focused on a question that keeps getting more urgent: can we trust what AI agents actually do?
Most of my recent work makes agent behavior measurable, and I ship all of it, across Python, TypeScript, and Rust. I also benchmark my own tools against their alternatives. When the result is unflattering, it goes in the README anyway.
- ratchetlock — regression testing for LLM prompts. Freeze outputs that pass, replay them in CI with zero LLM calls — built after a prompt fix quietly un-fixed itself four days in a row
- mycelium — hybrid RAG for markdown vaults, 98.0% Hit@5 on KorQuAD. Recently benchmarked it against agentic search and stock plugin embeddings
- adoptscore — reads conversation logs to find out which AI tools you actually use. Its own first headline (78% unused) failed an audit; the honest number is 38.6%
- daily briefing — turns each morning's AI news into Instagram cards and ships top-5 news Reels twice a day, no human in the loop (@todays.ai.brief, @muleori.news), scheduled with Airflow
| Project | What it does | Stack |
|---|---|---|
| mycelium | Local hybrid RAG + GraphRAG over any markdown vault. Korean-aware BM25 fused with dense retrieval reaches 98.0% Hit@5 on KorQuAD, with a public reproduction script. Answers cite their sources, and when retrieval confidence is low it says so instead of guessing. | Python |
| stateful-guardrails | Catches slow-burn, multi-turn crises by accumulating risk across the conversation instead of judging each message alone. Beats a 5-turn window by 38.0%p in crisis recall at 1/45 the cost of re-reading everything. Known weak spots are in the README. | Python |
| ratchetlock | Regression testing for LLM prompts, layered on promptfoo (its only runtime dependency). Freezes passing outputs as snapshots and replays them in CI with zero LLM calls, so a failure caught once can never quietly come back. Proven on a production card-news prompt: 0/5 → 4/5, with the remaining failure kept visible as a real catch. | TypeScript |
| agentscore | Lighthouse for AI agent dev environments — a CLI that scores the health of Claude Code MCP / plugin setups. | Python |
| adoptscore | Reads conversation logs to measure which AI coding tools you actually use, not which ones you installed. Audited its own 78% headline down to 38.6% and published the process. | Python |
| vali | Linter that flags hallucinations, slop, and over-engineering in AI-generated code. | TypeScript |
| Project | What it does | Stack |
|---|---|---|
| evmscope | EVM intelligence toolkit for AI agents — 23 MCP tools across 7 chains, zero config. | TypeScript |
| chain-eye | Real-time EVM transaction monitoring TUI. | Rust |
| Project | What it does | Stack |
|---|---|---|
| drift | Dependency health monitor — scores your project's dependency survival probability. | Rust |
| git-vibe | Vibe-check your codebase — git history analysis with emoji-based reports. | Rust |
| dotfig | One config to rule them all — generate ESLint, Prettier, TypeScript & EditorConfig from a single dotfig.yml. |
TypeScript |