DEVELOPER / BUILDER / LEARNER

Less magic.
More runtime.

I build the boring parts that keep AI agents alive —
permissions, recovery, and evals.

RUNTIME / TOOLS / RECOVERY / EVALS / RELIABLE SYSTEMS

01 / WORK

Selected work.

Systems for the moments
when software must hold.

01 / FEATURED PROJECT

AgentPermit4j

Guardrails for the part of an agent that can actually break things.

A trusted action execution layer for Java agents. It places authorization, risk policy, exact-call approval, idempotency, and audit between a proposed action and the code that executes it.

Explore the project

THE ACTION BOUNDARY

01Proposal
02Policy
03Approval
04Execute
05Audit

CONTROL BEFORE CONSEQUENCE

JAVA 21 / SPRING AI / POLICY / APPROVAL

02 / EXPERIMENTAL

RAGContract4j

Make bad retrieval fail CI, not production.

Turns retrieval, authorization, and citation requirements into repeatable quality gates.

↗

03 / TOOL

Claude Recovery Hooks

Claude died.
The task didn't have to.

PowerShell hooks for recovering Claude Code sessions after transient upstream failures.

↗

02 / NOTES

Field notes.

Writing on agents,
systems, and what lasts.

30 AUG 2026Agent 时代下的锚点从 Java 后端与企业知识检索实践出发，梳理 Agent 从 Demo 走向生产环境时不会轻易过时的八类工程能力。