AI Agents · Knowledge Systems · Codex Skills · Document Intelligence · Photography / Film
I like building small, practical systems that help people capture context, understand information, and turn messy inputs into useful outputs.
Current:
- 🧠 Agentic workflows — orchestrators, subagents, memory, tools, and diagnostics
- 📚 Knowledge tooling — course notes, second brain workflows, timeline evidence
- 📄 Document intelligence — PDF processing, layout parsing, OCR-friendly pipelines
- 🎞️ Visual notes — photography, vlog, and everyday observation
| Project | What it does | Stack / Direction |
|---|---|---|
| AgentWeave | A composable agent runtime with an orchestrator, isolated worker subagents, memory, artifacts, and diagnostics. | Python · Streamlit · OpenAI Agents SDK · Text2SQL · RAG |
| LLM-Deploy-Bench | A local React + Node/TypeScript console for acceptance and load testing LiteLLM model deployments. | TypeScript · React · Node.js · LiteLLM |
| dayflow-source | A Codex skill that turns local Dayflow timeline cards into compact evidence packages for reviews and summaries. | Python · SQLite · Codex Skill |
| course-media-to-notes | A Codex skill for organizing course videos, audio, PDFs, images, and subtitles into structured Chinese course notes. | Python · FFmpeg · ASR · PDF |
| dify_api_hub | A FastAPI support service for AI Agents, currently focused on PDF processing and layout parsing. | Python · FastAPI · PDF · Layout Parsing |
| obsidian-sync-yuque | Legacy Obsidian ↔ Yuque sync plugin. | TypeScript · Obsidian Plugin |
Make tools small enough to understand, useful enough to keep, and quiet enough to trust.
I care about:
- Local-first and privacy-aware workflows
- Clear evidence before generated summaries
- Composable agents instead of monolithic assistants
- Readable documentation and reproducible demos
- Tools that fit real daily work
When I am not building tools, I am probably framing things through a camera.