Python & AI Engineer · LLM / RAG / AI Agents · Quant Trading Systems
Indie builder at 喵哩文化創意 · Taiwan (UTC+8) · Open to remote roles
- 🛠️ I build and run production Python systems end to end: data in, decisions out, with monitoring on top.
- 📈 3 years as a quantitative trading researcher at a crypto hedge fund, then 3 years shipping my own trading infrastructure, LLM tools and an iOS app.
- 🔬 I care about honest validation: preregistered research, walk-forward tests and point-in-time universes. I also write post-mortems when the backtest turns out to be wrong.
- 🤖 I use Claude Code and Cursor every day, with tests and evidence-level "done" contracts keeping AI-assisted work safe.
| Project | What it is | Stack |
|---|---|---|
| AS Grid Dragon | Multi-exchange grid trading bot (Binance · Bybit · Bitget · Gate.io) built on an exchange adapter pattern, with WebSocket market data, backtesting and risk controls | Python · asyncio · FastAPI · Docker |
| Harmonic Pattern Scanner (private) | Telegram signal service for Binance USDT perpetuals: pattern detection, indicator confirmation, signal lifecycle tracking, 800+ automated tests, and a preregistered research program that removed survivorship bias | Python · ccxt · PostgreSQL · Alembic · Sentry |
| Alpha research pipeline (private) | Automated generator → simulator → gate audit → curator loop run by Claude Code subagents; I cut a core compile step from 127s to 31s | Python · SQLite · Claude Code |
| Multi-agent AI writing system (private) | Research, fact-check, edit and SEO agents that publish to WordPress; the v7 cleanup removed 16k lines and cut per-session context by 58% | Claude Code · MCP · Playwright |
| MeowDaigou | iOS pricing calculator for cross-border shopping agents; hit Top 3 in App Store Business on day three | SwiftUI · SwiftData · StoreKit 2 |
Languages
AI / LLM
Backend & Data
Ops
Quant Backtesting · Walk-forward validation · Reality Check / SPA · Monte Carlo risk models · ccxt
- Why a 92.6% backtest win rate didn't survive a point-in-time universe (survivorship bias)
- Win rate is a free parameter, so judge strategies by net expectancy
- A single factor beat my 8-factor model
- Evidence levels (D0–D4) for AI coding agents
Remote teams working on AI, LLM products, trading infrastructure or backend Python: let's talk.