Specimen 01 / 03

Taobao ChatBI

Ask a question about 10 million rows of Taobao behaviour logs in plain Chinese. A LangGraph workflow writes the SQL, checks it, runs it read-only, then explains the result.

• Metric definitions lifted trap-question accuracy from 70% to 100%
• A covering index cut a daily-active query from 30.5 s to 0.85 s
• Attack tests found and fixed a same-name CTE bypass of the SQL whitelist

• LangGraph
• LangChain
• DeepSeek
• MySQL / TiDB
• Streamlit

The live demo is password-protected to limit cost. Ask for access.

Specimen 02 / 03

AI Intraday Quant Trading Simulator

Four strategy families backtested on minute-level data for AAPL, TSLA, NVDA and CBA.AX, validated with purged walk-forward splits so nothing leaks from the future.

• About 55% out-of-sample accuracy still did not make money
• Judged by Sharpe, drawdown, win rate and payoff, not accuracy alone
• Backtests fill on the next candle with volatility-adjusted stops

• Python
• pandas
• scikit-learn
• XGBoost
• Streamlit

An educational simulator, not financial advice. The app sleeps when idle, so the first load can take a minute.

Specimen 03 / 03

Little Places

Ten small buildings you can look inside, in five colourways, each with three or four tiny interactions. Built on my own, with Claude Code, Codex and Three.js.

• Recurring mistakes turned into 30+ rules and a downloadable Claude Skill
• With the Skill, a more complex building took about half the tokens
• Auto-tour, Skill download page and a gallery, all live

• Three.js
• Claude Code
• Codex
• Claude Skill