实验室
views
小型、可检查的工程实验。每一项都来自已经实现的项目；没有独立部署时，会明确链接到源代码，而不是伪装成在线 Demo。
Experiments
Multi-agent protocol
Evidence-led agent collaboration
A working collaboration protocol that lets agents debate, retain evidence and reach a vote over a real-time stream.
- Next.js
- TypeScript
- SSE
- OpenAI
Agent × Web3
Safe, agent-callable Monad actions
A capability layer for discovery, simulation and unsigned transaction verification on Monad.
- TypeScript
- Monad
- MCP
- Web3
Agent infrastructure
Writer–Critic execution loops
A loop-driven execution harness that explores how specialized agents can produce, critique and refine work without hiding state.
- TypeScript
- Zod
- MCP
- CLI/TUI
LLM systems
A small end-to-end LLM pipeline
A study implementation spanning tokenizer training, pretraining, supervised fine-tuning and reinforcement learning.
- Python
- PyTorch
- BPE
- LLM
Interaction study
Three-dimensional tarot interactions
A shipped Web3 tarot prototype combining tactile 3D cards, prediction flows and Monad Testnet rewards.
- Solidity
- React
- Three.js
- Hardhat