Interdisciplinary AI builder working across agent systems, knowledge workflows, and applied ML portfolios.
name: "Zhongyue Lin (Leo) | 林中岳"
location: "China"
working_on:
- "Codex / Claude Code skill ecosystems"
- "Multi-model agent teams and AI workflow infrastructure"
- "Knowledge Hub: Obsidian-based course notes, research notes, and WeChat archives"
- "Applied ML and data-analysis portfolio projects"
research_bridge:
- "History and institutional design -> multi-agent orchestration"
- "Behavioral economics -> agent routing and evaluation"
- "Knowledge management -> durable AI memory workflows"
languages: ["中文", "English"]I build practical AI tooling around a simple question:
How do we turn messy human workflows into reliable, inspectable, reusable agent systems?
我关注的是把真实工作流沉淀成可复用的智能体系统：能调度模型、能整理知识、能留下上下文，也能在需要时被人工审查。
| Area | Canonical Repository | Status | Notes |
|---|---|---|---|
| Skills ecosystem | claude-code-skills | Active | Canonical public monorepo for installable Codex / Claude Code skills. 67 canonical installable units, 89 default install units. |
| Document parsing skill | mineru-skill | Active | Standalone MinerU API skill with CLI helper, examples, and release surface. |
| R analytics skill | r-analytics-skill | Active | R analytics pack with package-level data, stats, ML, spatial, NLP, and visualization workflows. |
| WeChat archive skill | wechat-public-account-archive-skill |
Private active | Archives WeChat public account articles into Knowledge Hub while keeping auth material out of notes. |
| YouTube clipping skill | youtube-clipper-skill |
Private active | Codex-ready video clipping workflow: download, semantic chaptering, bilingual subtitles, and summaries. |
| Knowledge base | Knowledge-Hub | Active | Obsidian-based CS self-study vault and long-term knowledge system. |
| Applied ML portfolio | udacity-masters-ai-projects | Active index | Portfolio entry point for Udacity AI / ML / data analysis projects. |
| Landmark classifier | landmark-classifier | Completed | M5 retraining artifacts merged; transfer model passes rubric and scratch CNN now has a stronger trained baseline. |
| Multi-agent research | civagent | Active / public | Historical governance systems as agent orchestration patterns. |
| macOS usage monitor | CodexBar | Fork / contribution | Usage stats for Codex and Claude Code, plus provider experiments. |
The skill ecosystem is being consolidated around one public canonical repository:
productivity Obsidian, NotebookLM, Google Workspace, Knowledge Hub
development MCP, frontend, repo design, D3, R analytics, SQL workflows
documents PDF, DOCX, PPTX, XLSX, MinerU, document coauthoring
research scientific skills, STEM modeling, history notes, macro research
ai-collaboration multi-provider orchestration, agent teams, skill discovery
product-management Lenny-inspired PM frameworks and playbooks
design themes, CLI demos, visual assets
system proxy, CDP, SVN, internal comms, OSS contribution
Recent consolidation work:
- Category toolkit routers now install correctly.
- Legacy wrappers such as
all-plan, provider-specific shims, andpm-*aliases are separated from the canonical install surface. - External nested submodule packs are documented without being flattened into the default install path.
- Older source repos are being treated as migration sources or standalone release surfaces, not as parallel canonical homes.
Standalone skills kept separate because they still have independent value:
| Repo | Why It Stays Separate |
|---|---|
| mineru-skill | Has CLI helpers, examples, changelog, and an independent document parsing workflow. |
| r-analytics-skill | Large package-level R workflow surface, useful as its own canonical pack. |
wechat-public-account-archive-skill |
Private local workflow with auth-sensitive operational scripts. |
youtube-clipper-skill |
Private media workflow with local configuration and output paths. |
I work on patterns for routing tasks across multiple model families and agent roles:
- role-based agent teams
- model/provider aliases
- smart task routing by skill, language, and intent
- inter-agent messaging and review loops
- durable context handoff between humans and models
Related repositories:
| Repository | Role |
|---|---|
| mmteam-cc | Multi-model agent team experiments for Claude Code-style workflows. |
| civagent | Governance systems as agent orchestration patterns. |
| xiaotie | Lightweight AI agent framework experiments. |
| Hivemind | Multi-AI routing and desktop collaboration prototype. |
| blog | Essays connecting history, social science, and agent architecture. |
My favorite ideas come from treating old institutions as tested coordination systems.
| Discipline | Source Pattern | AI / Agent Application |
|---|---|---|
| History | Tang three-department review, Persian satrapies, Athenian councils | quality gates, federated delegation, distributed judgment |
| Political science | governance modes and institutional constraints | agent topologies, authority boundaries, escalation paths |
| Behavioral economics | fast/slow thinking, bounded rationality | model routing, evaluator design, failure-mode analysis |
| Knowledge management | notes, links, indexes, evidence trails | durable memory, research archives, retrieval-ready context |
Blog entry point: leolin990405.github.io/blog
Main index: udacity-masters-ai-projects
Selected projects:
| Project | Focus |
|---|---|
| landmark-classifier | PyTorch landmark classification, MPS retraining, transfer learning, full submission artifacts. |
| teaching-llm-to-count | GRPO + LoRA experiments for teaching an LLM to count. |
| agentsville-trip-planner | Agentic AI trip planning with ReACT-style workflows. |
| face-generation | DCGAN face generation project. |
| text-translation-sentiment | Translation + sentiment pipeline using Hugging Face tooling. |
| sentimentscope-imdb | Transformer-based IMDB sentiment analysis. |
Older course repositories remain useful as project records, while the portfolio index is the preferred entry point.
I am gradually reducing repo sprawl:
| Keep Maintaining | Consolidate / Redirect | Archive When Safe |
|---|---|---|
claude-code-skills |
claude-lenny-skills -> product-management packs |
old migrated skill source repos |
mineru-skill |
claude-office-skills -> documents packs |
thin wrappers with no independent release surface |
r-analytics-skill |
claude-design-skills -> design packs |
stale forks without active PRs |
Knowledge-Hub |
claude-obsidian-skills -> productivity packs |
one-off course repos after portfolio indexing |
civagent |
github-repo-design-skill -> development pack |
duplicate awesome lists after blog/Knowledge Hub import |
Principles:
- one canonical home per workflow
- no secrets or auth material in public repos
- private repos for operational scripts with local credentials
- public repos for reusable frameworks, skills, and portfolio artifacts
- README files should tell users where the maintained surface now lives
- GitHub: @LeoLin990405
- Blog: leolin990405.github.io/blog
- LinkedIn: Zhongyue Lin
If one of my projects is useful, a star is always appreciated.