I'm teaching myself to build AI products — by building, in public.
Right now I'm building facta, a personal AI agent written from scratch: LLM loop, RAG over my own notes, tool calling, persistent memory. No frameworks — the goal is to understand every layer before shipping anything on top of it.
- LLM agents: orchestration, memory, retrieval
- Local-first software
- Turning side projects into real products
- 🔨 Building: facta — a from-scratch personal AI agent (now open source): ReAct loop, layered memory, RAG knowledge layer, 9 built-in tool families + MCP, crash self-recovery, CLI + Web UI
- 📖 Learning in public: 12-week systematic AI curriculum — ML foundations → deep learning → CNN/RNN/Transformer (day 38/84, in progress)
- 🔬 Weekly deep-dives: one high-quality repo or paper per week
- facta — a personal AI agent built from scratch: ReAct loop, layered memory, RAG knowledge layer (BGE-M3 + ChromaDB + knowledge graph), 9 built-in tool families + MCP client, crash self-recovery, safety gates, CLI + Web UI. No agent frameworks — every layer hand-written and replaceable
- learning-notes — learning in public（中文笔记）: a 12-week AI study series with full derivations, code and self-checklists, plus weekly repo deep-dives and paper write-ups
- md-reader — local-first Markdown reader, packagable as a macOS desktop app (React + Electron)
| 日期 | 内容 |
|---|---|
| 2026-09-28 | AI 系统学习笔记 · 第 1–39 天（ML 基础 / 损失函数 / 优化 / CNN，推导 + 代码 + 自测） |
| 2026-09-25 | vLLM：用分页思想管好 KV Cache |
| 2026-09-25 | Agent 的"决定"，有 20.4% 在执行时悄悄变了样 |