Fortis Fortuna Adiuvat. I'm CachoidXie — an independent researcher and developer in Hangzhou working on local-first agent safety, competitive-programming infrastructure, and language-model research.
An online-judge platform built for schools, designed to connect classroom teaching with AI. The AI gateway that grades, hints and explains submissions is still in active development.
A local-first Agent Harness Kernel with Git-like control for safe, auditable AI agents. Capability capsules, an event-sourced memory OS, and a policy engine — so every action an agent takes is reviewable.
PyTorch research scaffold for RetNet, block attention residuals, and hashed Engram memory for long-context language modeling.
A brand-new shortest-path algorithm derived from the latest research out of Tsinghua University.
A kernel that treats decisions like Git history turns "trust the AI" into "read the diff." Notes on capability capsules and event sourcing.
ON LONG CONTEXTHashed memory traces, retention curves, and why RetNet's recurrence makes retrieval cheaper than you'd expect at scale.
I work as an independent researcher and developer. My work lives at the intersection of AI safety, competitive-programming infrastructure, and language-model research. I build because I want systems that are legible — you should be able to read what a machine decided and why.
Shipping the AI gateway for a school-grade online judge in Rust.
RetNet, block attention residuals, and hashed Engram memory in PyTorch.
Local-first agent harnessing with Git-like auditability and a policy engine.
Collaborations on agent safety, OJ systems, and ML infrastructure.