Building agent
runtimes, judges,
and research.
与机器智能共事，做可被审计的系统。
I build local-first agent infrastructure, AI online judges, and language-model research scaffolds. Systems you can audit because you can read every decision they make.
Selected work
Five projects spanning agent accountability, local-first harnesses, AI online judging, and long-context language modeling.
Agnet
An early version of Internet-of-Agent. We need accountability in the era of agents - a protocol layer (a2a) where every agentic action carries a traceable accountability chain.
Aetherion
Local-first agent harness kernel with Git-like control. Capability capsules, event-sourced memory, policy engine for safe agents.
CodeNexus
An online judge for schools, connecting teaching with AI. The grading and hinting gateway is in active development.
Anamnesis
PyTorch scaffold for RetNet, block attention residuals, and hashed Engram memory for long-context modeling.
Shortest-Path
A brand-new shortest-path algorithm from the latest research out of Tsinghua University.
Accountability for the agent era.
My work converges on one question: as agents act on our behalf, how do we keep every action legible, reversible, and attributable? Three threads run through it.
Agent accountability chains
Protocol-level traces so every agentic action carries provenance, not just output.
Local-first harnessing
Capability capsules and event sourcing. Git-like auditability for agent runs.
Long-context memory
RetNet recurrence and hashed Engram memory. Cheaper retrieval at scale.
Working on legible machine intelligence.
I am an independent researcher and developer based in Hangzhou. My work sits at the intersection of AI safety, competitive-programming infrastructure, and language-model research.
I build because I want systems you can read. Not black boxes that decide for you, but agents whose every action you can trace, question, and reverse. That belief runs through Agnet, Aetherion, and Anamnesis alike.
Agnet - Internet-of-Agent
Defining the accountability chain for inter-agent protocols.
Anamnesis
RetNet, block attention residuals, hashed Engram memory.
Collaboration
Agent safety, OJ systems, and ML infrastructure.
Let's build
something legible.
Working on agent safety, an online judge, or research infrastructure? Open to collaboration and conversation.