Overview

Full-stack developer based in Taiwan, building personal AI agents, vertical tools for the Claude ecosystem, and production systems across robotics and quantitative finance.

My work spans three tracks: 🦞 Claw Ecosystem — a personal AI agent product line delivered through messaging apps with MCP integration (NanoGemClaw); 🤖 Claude Ecosystem — vertical plugins and skills that extend Claude Code and Claude cowork into specialized workflows (Claudistotle for philosophy research, ai-agent-architect for AI agent architecture design); and 🎯 Domain Applications — real-world systems solving problems in robotics (ROSForge) and Taiwan equity valuation (TaiwanStockVECalculator). When I'm not shipping products, I deep-dive into high-impact open-source AI projects — analyzing architectures, adding benchmarking toolkits, and building optimization layers.

---

Flagship Projects

1. NanoGemClaw — Gemini-Powered Google Ecosystem AI Assistant

🦞 Claw Ecosystem · Personal AI Agent
A modular AI assistant on Telegram with MCP Client Bridge, smart routing, and deep Google ecosystem integration.

What it does: A full-featured AI assistant powered by Gemini, delivered via Telegram with a 12-page real-time web dashboard. 7 built-in plugins (including Google ecosystem) turn it into a unified personal productivity hub — search Drive files, manage Calendar events, sync Tasks, and query a hybrid RAG knowledge base, all through natural conversation. Key differentiators vs NanoClaw (Claude-based): NanoClaw NanoGemClaw Agent Claude SDK Gemini + MCP Client Bridge Bot Framework node-telegram-bot-api grammY (type-safe) Messaging WhatsApp Telegram Bot API Cost $100/mo Free tier (60 req/min) Architecture Monolith Modular monorepo (8 packages + 7 plugins) Media Text only Photo, Voice, Video, Document Google Integration None Drive, Calendar, Tasks, RAG Tests ~50 1,244 tests (54 files, 92% coverage) Highlights: grammY Bot Framework — Type-safe, event-driven Telegram integration with rate limiting and message consolidation MCP Client Bridge — Model Context Protocol support with per-tool whitelist and Zod schema validation Smart Routing — preferredPath intelligently switches between fast path (direct API) and container execution Google Ecosystem — 7 plugins: OAuth hub, Drive, Calendar (full CRUD), Tasks (bidirectional sync), Drive Knowledge RAG (hybrid search), Discord Reporter, Memorization Service Plugin System — 6 extension points: Gemini Tools, Message Hooks, Express Routes, IPC Handlers, Background Services, Dashboard Extensions Fast Path — Direct Gemini API streaming with context caching (75-90% token cost reduction) and native function calling Hybrid Drive RAG — Pre-indexed embeddings via physical file approach + live Drive search fallback 12-page Dashboard — Overview, Tasks, Calendar, Drive Browser, Knowledge, Analytics, Memory, Logs, Activity, Settings, Schedule, MCP Management i18n 100% — Full interface support for 8 languages 1,244 tests across 54 files — 92% statement coverage, 84% branch coverage

Monorepo nanogemclaw/ ├── packages/ │ ├── core/ │ ├── db/ │ ├── gemini/ │ ├── telegram/ │ ├── server/ │ ├── plugin-api/ │ ├── event-bus/ │ └── dashboard/ ├── plugins/ │ ├── google-auth/ │ ├── google-drive/ │ ├── google-tasks/ │ ├── google-calendar-rw/ │ ├── drive-knowledge-rag/ │ ├── discord-reporter/ │ └── memorization-service/ ├── app/ ├── container/ └── docs/ Google Tools (16) Drive → search, read, summarize Tasks → create, complete, list Calendar → create, list, update, delete, check_availability RAG → search_knowledge Discord → daily/weekly reports

git clone https://github.com/Rlin1027/NanoGemClaw.git
cp .env.example .env    # Add TELEGRAM_BOT_TOKEN + GEMINI_API_KEY
npm install && npm run dev

---

2. Claudistotle — Philosophy Research Plugin for Claude Code

🤖 Claude Ecosystem · Academic Research Pipeline
A full-pipeline philosophy research assistant: from topic idea to peer-reviewed paper, entirely inside Claude Code.

What it does: Claudistotle takes you end-to-end from a rough research question to a polished academic paper — searching 8+ academic databases, synthesizing a literature review with verified citations (never fabricates references), writing structured academic prose, and simulating three-round peer review. Built as a Claude Code plugin. Pipeline: Topic idea → Research design → Literature review → Draft → Peer review → Paper Highlights: 8+ academic databases — Semantic Scholar, OpenAlex, CORE, arXiv, PhilPapers, Stanford Encyclopedia, Internet Encyclopedia, Notre Dame Philosophical Reviews Citation integrity — every reference verified against CrossRef; fabricated citations are impossible by design (BibTeX hooks validate metadata provenance before any .bib is written) 6-phase literature review — environment check → domain decomposition → parallel database search → synthesis planning → section writing → assembly with DOCX export Three-round peer review simulation — two AI personas (Athena the reviewer, Calliope the reviser) iterate on argument validity, evidence quality, and writing clarity Autopilot with self-healing — cautious / moderate / full autonomy modes; automatic retry on quality gate failures Text commentary — close reading and logical annotation of primary philosophical texts, auto-integrated into your draft Resumable — interrupted sessions resume from PROGRESS.md Extends PhilLit — the multi-agent research workflow from Syracuse University & University of Hamburg

Commands (12) setup help research-design literature-review text-commentary draft peer-review refine validate feedback autopilot philosophy-research Output reviews/[project]/ ├── research-proposal.md ├── literature-review-final.md ├── literature-all.bib ├── argument-skeleton.md ├── paper-draft.md ├── change-record.md ├── PROGRESS.md └── sources/ ├── primary/ └── secondary/

# Install
/plugin marketplace add Rlin1027/claudistotle
/plugin install claudistotle@claudistotle

# Use
/claudistotle:setup
/claudistotle:literature-review

---

3. ai-agent-architect — AI Agent Architecture Designer

🤖 Claude Ecosystem · Socratic Agent Design Pipeline
A Claude cowork skill that interviews you before recommending any AI agent architecture — grounding every design decision in actual requirements.

Problem it solves: Most AI agent failures come from choosing patterns based on technical appeal rather than real business requirements. Teams jump to "multi-agent system" before understanding what problem they're solving. Approach: A structured five-phase pipeline based on Anthropic's Building Effective Agents framework. It runs a Socratic deep interview across four clarity dimensions, scores ambiguity in real time, and only produces a recommendation once the problem is actually understood. Highlights: Five-phase pipeline — Deep Interview → Recommendation → Blueprint → Functional Modules → Execution Bridge, each with a quality gate Four-dimension scoring — Goal (35%) / Constraints (25%) / Domain (25%) / Success Criteria (15%), live progress bars every round Ambiguity gate — pipeline won't exit Phase 1 until ambiguity ≤ 20% (soft cap Round 8, hard cap Round 12) Challenge modes — auto-triggered Contrarian / Simplifier / Ontologist questions break circular answers and surface hidden assumptions Decision matrix — four-axis evaluation (control × complexity × resources × expertise) selects from 8 architecture patterns including hybrid strategies Single-agent first — multi-agent systems (10–15× more tokens) only recommended when interview evidence clearly shows a single agent is insufficient Greenfield & brownfield — automatically shifts questioning when you mention migration, legacy integration, or rollback criteria Implementation-ready output — blueprint with Mermaid architecture diagram, PRD, technical spec, cost analysis spreadsheet, risk matrix, evaluation plan, and starter prototype

Phases Phase 1 · Deep Interview └ gate: ambiguity ≤ 20% Phase 2 · Recommendation └ gate: user confirms Phase 3 · Blueprint Phase 4 · Functional Modules Phase 5 · Execution Bridge Patterns (8) Single Agent Routing Workflow Sequential Workflow Parallel Workflow Hierarchical Multi-Agent Collaborative Multi-Agent Evaluator-Optimizer Hybrid Strategies Artifacts interview-spec.md ai-agent-blueprint.md prd.md tech-spec.md cost-analysis.xlsx risk-matrix.md eval-plan.md prototype/

git clone https://github.com/Rlin1027/ai-agent-architect.git
# Place into your Claude cowork skills directory
# Trigger: mention "AI agent design" or "agent architecture" in your prompt

---

Domain Applications

ROSForge — AI-Powered ROS1 to ROS2 Migration Engine

🎯 Domain Applications · Robotics
The first AI-driven tool to automate legacy robotics code migration.

Problem: Migrating a mid-size ROS1 package (~5K-10K LoC) takes a senior engineer 2-4 weeks of manual refactoring. With ROS1 Noetic EOL (May 2025), thousands of packages face abandonment. Solution: One command to reforge your legacy robotics packages. ROSForge uses LLMs to understand code semantics — not just regex-replace — and delivers end-to-end migration with automatic build verification and fix loops. What makes it unique: BYOM (Bring Your Own Model) — freely switch between Claude, Gemini, and OpenAI Full-spectrum migration — C++ (roscpp→rclcpp), Python (rospy→rclpy), Launch files (XML→Python), CMakeLists.txt, package.xml, msg/srv definitions Validate-Fix loop — auto colcon build + AI-driven error repair Interactive mode — pause at critical steps for human review Workspace-level batch migration with cross-package dependency resolution Custom rules via .rosforge/rules.yaml

Stack Python 3.14 | Click CLI Pipeline Architecture ├─ Ingest (Parsers) ├─ Analyze (Deps) ├─ Transform (AI) ├─ Validate (Build) └─ Report (Diff) Pipeline rosforge migrate ./pkg ↓ Parse AST/IR ↓ Resolve dependencies ↓ AI transform (BYOM) ↓ colcon build ↓ Auto-fix if failed ↓ Generate report

pip install rosforge
rosforge migrate ./my_ros1_package          # End-to-end migration
rosforge analyze ./my_ros1_package          # Analysis only (no changes)
rosforge config set engine claude-code      # Switch AI engine

---

TaiwanStockVECalculator — 7-Model Stock Valuation System

🎯 Domain Applications · Quantitative Finance
Quantitative valuation engine for Taiwan stocks with LLM-enhanced classification and adaptive feedback loops.

What it does: Input a stock ticker, get fair value estimates from 7 independent models — weighted by industry classification, validated by backtesting, and auto-adjusted through feedback loops. 7 Valuation Models: Model Method A Multi-stage DCF (Discounted Cash Flow) B Dividend Yield + Gordon Growth Model C PER (Price-to-Earnings Ratio) River Chart D PBR (Price-to-Book Ratio) E CapEx Forward-Looking Valuation F EV/EBITDA Enterprise Value Multiples G PSR (Price-to-Sales Ratio) System features: 3 interfaces — CLI (colored terminal), HTTP API (Express), n8n automation workflow LLM smart classification — auto-categorize stocks by industry with guardrails Backtest engine — 90d/180d accuracy validation with hit rate, MAE metrics Adaptive weights — feedback loop adjusts model weights based on backtest accuracy per category Portfolio management — sector allocation, risk metrics, performance tracking Alert system — price/valuation/classification change triggers → Telegram/Email push

Architecture src/ ├── api/ # FinMind API ├── models/ # 7 valuation models │ ├── dcf.js │ ├── dividend.js │ ├── per.js │ ├── pbr.js │ ├── capex.js │ ├── ev-ebitda.js │ └── psr.js ├── llm/ # LLM guardrails ├── backtest/ # Accuracy engine ├── feedback/ # Adaptive weights ├── portfolio/ # Analytics └── report/ # Synthesis Data Pipeline FinMind API (8 datasets) ↓ Parallel fetch 7 Models (parallel calc) ↓ Smart weighting LLM Classification ↓ Backtest feedback Final Recommendation ↓ n8n automation Telegram / Email

node src/index.js 2330              # CLI: Analyze TSMC
node src/server.js                  # Start HTTP API
curl -X POST localhost:3000/api/analyze/2330

---

Tech Stack

---

Other Projects

AI & Developer Tools

Project	Description	Stack
repo2prompt	Convert git repos to LLM-friendly prompts with token estimation	Python
pocketflow-enhanced	Visualization, tracing & caching extensions for the 100-line LLM framework	Python

Quantitative Finance & Full-Stack

Project	Description	Stack
heritage-hunter	Gamified unclaimed land search engine for Taiwan with interactive map	Next.js, Supabase
fire-calculator	Financial Independence / Retire Early calculator	TypeScript
n8n-portfolio	Production-grade automation workflows with AI, RAG & data pipelines	n8n, Python
SLOTprototype	Fortune God Slots — Chinese themed slot machine prototype	TypeScript

---

AI/ML Research Toolkit

I systematically enhance popular open-source AI projects by adding architecture analysis, benchmarking frameworks, and optimization toolkits.

25+ projects enhanced across LLMs, TTS, Video, Agents, and Training Infrastructure — totaling 3,000+ tests written.

LLM & Language Models — 8 projects

Project	Original	What's Added
nanochat-enhanced	Karpathy's nanoGPT (43k stars)	Architecture analysis, optimization toolkit & benchmarking — 167 tests
nano-vllm-enhanced	nano-vllm	Analytics, advanced sampling & optimization — 160 tests
grpo-zero-enhanced	DeepSeek R1 GRPO	Analytics, Math24/logic tasks, algorithm variants — 105 tests
hrm-enhanced	Hierarchical Reasoning Model	Analytics, puzzle generation & advanced algorithms — 147 tests
lingua-enhanced	Meta Lingua	Architecture analysis, config library & training estimation — 137 tests
minbpe-enhanced	Karpathy's minbpe	WordPiece/Unigram/BPE-Dropout algorithms & visualization — 96 tests
picoGPT-enhanced	picoGPT	Sampling, KV-Cache & interactive mode for GPT-2 in NumPy
tiny-llm-enhanced	LLM Serving Course	Architecture analysis, serving strategy & benchmarking

AI Agents & MCP — 8 projects

Project	Original	What's Added
swarm-enhanced	OpenAI Swarm (21k stars)	Multi-agent analysis, optimization & benchmarking — 131 tests
aisuite-enhanced	Andrew Ng's aisuite (13.5k stars)	Provider analysis, intelligent routing & benchmarking
fastapi-mcp-enhanced	fastapi-mcp (11.5k stars)	Endpoint analysis, intelligent routing & conversion benchmarking
claude-agent-sdk-enhanced	Anthropic Agent SDK (4.8k stars)	Configuration analysis, optimization & benchmarking
langchain-mcp-enhanced	LangChain MCP Adapters	Configuration analysis, performance optimization — 211 tests
mcpo-enhanced	mcpo (MCP-to-OpenAPI)	Config analysis, intelligent routing & benchmarking
claude-usage-monitor-enhanced	Claude Code Usage Monitor	Usage analysis, cost optimization & benchmarking — 128 tests
simple-evals-enhanced	OpenAI simple-evals	Analysis, statistics & reporting for LLM evaluations

Vision & Video Generation — 5 projects

Project	Original	What's Added
framepack-enhanced	FramePack (16.6k stars)	Architecture analysis, optimization & benchmarking — 227 tests
omost-enhanced	Omost (7.6k stars)	Canvas analysis, pipeline optimization & benchmarking — 148 tests
ltx-video-enhanced	LTX-Video (9.3k stars)	Architecture analysis, optimization & benchmarking
vjepa2-enhanced	Meta V-JEPA 2	Architecture analysis, config library & benchmarking
mambaout-enhanced	MambaOut (CVPR 2025)	Architecture analysis, model variants & benchmarking — 113 tests

Speech & Audio — 4 projects

Project	Original	What's Added
dia-enhanced	Dia TTS (19.1k stars)	Architecture analysis, optimization & benchmarking — 172 tests
csm-enhanced	Sesame CSM (14.5k stars)	Architecture analysis, inference optimization & benchmarking
pocket-tts-enhanced	Kyutai Pocket TTS (3.1k stars)	Synthesis profiling & hardware benchmarking
speech-to-speech-enhanced	HuggingFace S2S	Pipeline analysis, config management & benchmarking

Training Infrastructure & Architecture — 5 projects

Project	Original	What's Added
self-forcing-enhanced	Self-Forcing (NeurIPS 2025 Spotlight)	Training analysis, pipeline optimization & benchmarking
dualpipe-enhanced	DualPipe (DeepSeek V3/R1)	Pipeline analysis, schedule simulation & benchmarking — 132 tests
nano-graphrag-enhanced	nano-graphrag (GraphRAG)	Graph analysis, retrieval strategies & benchmarking
efficient-kan-enhanced	efficient-kan (KAN)	Analysis, visualization & training utilities
minimalRL-enhanced	minimalRL (12 RL algorithms)	Logging, visualization & experiment tracking

---

GitHub Stats