English · 简体中文
I build the machinery beneath intelligent software: agent runtimes and context systems,
from cloud infrastructure to on-device intelligence.
I care about making probabilistic systems dependable without making their complexity the user's problem.
Agent Runtimes · Context Engineering · Cloud Architecture · On-Device Intelligence
I'm currently building Axiom Loom, a local-first agent runtime for continuous context, memory, and multi-agent orchestration.
The idea is simple: finite context windows should remain an implementation detail, not something users have to manage.
Rigorous underneath. Effortless in motion.
These projects explore the same idea at different scales: give code clear responsibilities, keep context deliberate, and make an agent's behavior inspectable.
| Project | What you can explore |
|---|---|
| The Regular | A resident QQ group companion with code-owned relationships, rolling memory and an offline demo. |
| Neural-Janitor | Preview-first tab cleanup for Chrome and Edge on macOS, with a local Swift companion that learns personal retention patterns. |
| Loop Warden | A Hermes skill and Python helper for bounded delegation to local coding CLIs, with explicit task contracts and structured handoff. |
| The Norn Machine | Deterministic context assembly behind an image-selection demo: Python builds a behavioral profile, and the model renders it in language. |
| AgenticX Decision | A document-review research harness with bounded retrieval, explicit Python rule tools and inspectable evidence. |
I tend to put deterministic control wherever authority matters, keep personal data local when I can, and design architectures that can outlive any one provider. I like useful defaults more than walls of configuration, and I try to back the claims that matter with measurements.
That usually leaves me working at the point where models stop being demos and start becoming dependable software.