I build small open-source tools for the parts of LLM apps that are easy to skip: output checks, evals, RAG and local models.
Lately
- Output and tool-call checks — schema shims, tool-call protocol, guardrail comparison
- Evals — LLM-judge agreement, embedding A/B tests, RAG recall
- Local LLM ops — Ollama, LoRA adapters, context windows
- Korean text for LLMs — PII masking, TTS readout, IME testing
Most projects run offline with a mock mode, no API key needed. Each comes with tests, CI and an MIT license.
한국어 소개
LLM 앱을 만들다 "이런 도구 하나 있으면 좋겠다" 싶던 것을 직접 만들어 오픈소스로 공개합니다. 요즘은 출력·도구 호출 검증, 평가, 로컬 LLM 운영, 한국어 처리를 주로 봅니다. 대부분 API 키 없이 모의 모드로 바로 돌려 볼 수 있습니다.
- align16 — Find which Flutter plugin ships a non-16KB-aligned .so
- langchain-v1-migrate — Migrate langchain 0.3 imports to 1.x with a computed move table
- mcp-prompt-commands — Turn MCP server prompts into slash commands and skills
- bases-lint — Lint Obsidian Bases files before they fail silently
- ckpt-doctor — Find the broken shard in a safetensors model folder
- emucheck — Preflight which AWS calls a local emulator can serve
- hubcard-lint — Lint Hugging Face repo card metadata before you push
- hwpx-doctor — Check and safely fix generated HWPX files in CI
All 119 projects: ai-lab
Python · TypeScript · Dart



