I build applied AI systems — agents, data tools, and LLM applications that hold up in production.
From replay bytes to player paths, teamfights, and match insights. A Python parser that turns Dota 2 .dem files into DataFrames, JSON, Parquet, and interactive HTML reports.
Player movement reconstructed from a replay. Open the sample report to explore.
Code → · Docs → · Sample report →
A coding agent with resumable sessions, approval controls, and inspectable transcripts. One shared runtime across terminal, RPC, and SDK interfaces.
Code changes, tool output, and session commands in the terminal.
| Project | What I'm building |
|---|---|
| Krill.jl |
A Julia-native AI agent runtime for Telegram and Discord, with persistent memory, scheduled tasks, MCP tools, and coding-agent delegation. In active development. |
| OpenCouch |
A pre-beta AI companion for self-reflection and guided wellness exercises, with persistent memory and safety routing. |
| QuantRL-Lab |
A Python testbed for reinforcement learning in quantitative trading, with pluggable action, observation, and reward strategies. |
ai-agent-book · contributor
One of the top contributors to Bojie Li's open-source book on AI agent design and engineering. My contributions span English translation improvements, chapter experiment organization, Python dependency consistency, and test refactoring.
Read the English edition → · My merged contributions →
- AI agents and LLM-powered applications
- Reinforcement learning and quantitative finance
- Rust and Julia for high-performance computing
- 2026-09-04 · gem-dota v0.7.1 — Improved full-replay parsing performance while preserving normalized output, and added reproducible replay fixtures for integration tests.
- 2026-08-23 · Wisp v0.1.0 — Released terminal steering and follow-up controls, reduced tool-output flicker, and streamlined startup.
- 2026-07-31 · ai-agent-book contribution — Merged shared provider infrastructure, chapter dependency setup, and experiment-structure improvements into the upstream book.