Xiang Ding

Projects / things I build

Selected projects

Things I build: agent systems and training pipelines for work and research, plus personal projects around games and characters.

Kirakira Agent banner: an empty concert stage with instruments under soft pastel lights.
AGENT RUNTIME

Kirakira Agent

A local-first personal assistant that chats, reaches out on its own, and keeps working in the background across Web, Telegram, and QQ, with long-term memory built from past conversations.

  • Python
  • ReAct
  • MCP
  • Long-term memory
GitHub
ShopSimulator overview: a shopper with a profile talks to an LLM agent, which searches and clicks through a shop environment before recommending a product.
AGENTIC RL

ShopAgent RL

Teaches Qwen3.5-4B to shop for a simulated customer — ask what they need, search, compare, and buy — with agentic SFT and then online GRPO. Strict success on 1,459 held-out tasks rises from 13.16% to 64.63%.

Environment figure from ShopSimulator (Alibaba), the simulator this project trains in.

  • Qwen3.5-4B
  • SFT
  • GRPO
  • veRL
GitHub
Shiori first-run screen: choose game folders on the left, connect a Bangumi token on the right.
DESKTOP APP

Shiori 栞

A Windows galgame launcher. Point it at your folders and it finds the games, launches each one the way it needs — Locale Emulator or a temporary virtual disc — and fills in details from Bangumi and VNDB.

  • Electron
  • Bangumi
  • VNDB
GitHub
Sora demo page: Niimi Sora waving beside a chat panel and expression controls.
INTERACTIVE DEMO

Sora 天色微风

A character study of Niimi Sora from 9-nine- that runs in the browser. Her eyes follow the cursor, she reacts to touch with a dozen expressions, and she can chat and talk back with AI-generated voice.

  • StandRig
  • LLM chat
  • TTS
Open demo GitHub

OPEN SOURCEOpen source & community

CONTRIBUTOR

PenguinHarness

An open-source framework where agents improve themselves round by round: the model stays fixed while prompts, Skills, and tools evolve. I contribute fixes to its RSI algorithm toolkits.

  • RSI
  • Agent harness
  • Skills
  • Benchmarks
COMMUNITY

Shared RSI Community

An open community co-building benchmarks, evaluations, and algorithms for recursive self-improvement.

  • Benchmark
  • Harness
  • Algorithm
  • Evaluation