Hi, I'm Wumen. Welcome to my workspace. 你好,我是無門。欢迎来到我的 workspace。

Have an idea?
Let AI finish it.
想到一件事,
就让 AI
把它做完。

無門的头像

Workspace

Screens from things I actually built 桌面上摊开的,都是真做过的东西
01 6 6

agent projects 个 Agent 项目

built in 2026, unfinished ones included 2026 年做的,含做到一半的
02 2 2

open-source repos 个开源仓库

AIOps agent, observability base AIOps Agent、可观测底座
03 4 4

real-world domains 个真实场景

ops, hiring, renting, healthcare 运维、招聘、租房、医疗
04 10 10

field notes 篇实践笔记

as of Oct 2026 快照 2026-10

About

About me 关于我

我 是 無 門 , 程 序 员 , 云 原 生 和 运 维 出 身 。 这 一 年 我 把 遇 到 的 真 问 题 一 件 件 交 给   A g e n t : 排 查 故 障 、 找 房 、 找 工 作 、 陪 外 国 人 看 病 。 有 的 开 源 了 , 有 的 做 到 一 半 停 了 , 都 摆 在 这 里 。 比 起 结 论 , 我 更 想 让 你 看 到 它 们 是 怎 么 被 拆 开 、 跑 起 来 、 被 检 查 的 。

I ' m   W u m e n ,   a   p r o g r a m m e r   f r o m   c l o u d - n a t i v e   a n d   o p s .   T h i s   y e a r   I   h a n d e d   r e a l   p r o b l e m s   t o   a g e n t s   o n e   b y   o n e :   i n c i d e n t   t r i a g e ,   r e n t i n g ,   h i r i n g ,   m e d i c a l   e s c o r t   f o r   f o r e i g n e r s .   S o m e   a r e   o p e n   s o u r c e ,   s o m e   s t o p p e d   h a l f w a y .   T h e y   a r e   a l l   h e r e ,   w i t h   h o w   t h e y   w e r e   t a k e n   a p a r t ,   r u n ,   a n d   c h e c k e d .

Works

Things I built 做过的东西

Each one is labelled honestly: open source, MVP, experiment, or paused halfway. 每个都标了实际状态:开源、MVP、实验,或者做到一半停了。

01
AI Agent / AIOps Open source 开源

AIOps 故障诊断 Agent

When a Pod breaks, the agent investigates and reports with evidence. Pod 出了问题,Agent 自己去查,带着证据写报告。

GitHub

For Kubernetes Pod incidents, the agent pulls status, logs, events, and metrics, follows runbooks to narrow things down, and writes a report where every conclusion carries its evidence. 面向 Kubernetes 的 Pod 异常。Agent 自动拉状态、日志、事件和指标,对照 Runbook 一步步缩小范围,最后交一份每条结论都挂着证据的诊断报告。

  • Collects status, logs, events, and metrics via tools 自动调用工具采集状态、日志、事件和指标
  • Reasons against runbooks into an evidence chain 对照 Runbook 推理,形成证据链
  • Writes structured reports, measured with an eval set 输出结构化诊断报告,配评估集测效果
02
AI Agent / 医疗服务 MVP MVP

MedBridge · 外籍患者陪诊

Helps foreigners in Beijing see a doctor: what to do, what it costs, and a person to go with them. 帮来北京看病的外国人搞清楚怎么看、要花多少钱,再有真人陪着去。

A bilingual chat agent up front and a fulfilment console behind it: quote, booking, visit-day updates, visit summary, and review, one order end to end. Chat and phase-one fulfilment are done; the escort roster is still a placeholder. 前台是中英双语的咨询 Agent,后台是给协调员用的履约系统:从报价、预约、就诊当天的节点通知,到就诊小结和评价,一单从头跑到尾。咨询和履约一期都做完了,陪诊师名单还是占位,没上线。

  • Collects needs in chat and quotes line by line 对话采集就医需求,当场给逐项透明报价
  • Ops console for booking, visit-day steps, and summaries 运营后台管预约、就诊当天节点和就诊小结
  • Collects reviews and strips private details before publishing 就诊后收评价,发布前自动去掉隐私信息
  • Bilingual, points emergencies to 120, gives no medical advice 中英双语,急症指引 120,不给医疗建议
03
AI 产品 / 招聘 Paused halfway 做到一半

無門職言 · AI 人才市场

Job posts that tell the truth, candidates who show real work. 让岗位说真话,让候选人用作品说话。

A candidate-first hiring market. Every job field shows whether its source is confirmed; candidates build profiles from work evidence; employers commit to review capacity and reply on time. The local loop works; billing and growth were not built. 候选人优先的招聘平台。岗位信息逐字段标出来源是否已确认;候选人用作品证据建档;企业承诺审核容量,按时限给真人回应。本地产品闭环已打通,结算和获客没做,停在上线前。

  • Field-by-field sourcing on every job post 岗位信息逐字段溯源,已确认和待确认分开标
  • Evidence-based profiles, shared only after the owner confirms 候选人用作品证据建档,本人确认后才对外
  • Same capabilities as web UI, MCP server, and CLI 同一套能力同时提供网页、MCP 和命令行
04
AI Agent / 生活 MVP MVP

安巢 · 北京租房 Agent

Say what you need; it ranks homes by real commute and shows the catches. 一句话说需求,按实测通勤给你挑房,坑也摆出来。

Give your budget, office, and commute limit; the agent filters listings by real map-API commute times and shows hits and risks for each. Runs end to end locally on a sample listing pool. 说出预算、公司位置和通勤底线,Agent 用高德实测通勤筛房源,每套房亮出命中项和风险提示。本地闭环跑通,房源用的是样例池,没接真实平台。

  • Plain-language needs, ranked by measured commute and evidence 一句话说需求,按实测通勤和证据排序
  • Every listing shows its hits and its risks 每套房列出命中项和风险,不藏坑
  • Remembers your taste, editable, with a one-click handoff 记住并可纠正你的偏好,一键生成顾问交接单

More experiments 还做过这些

多 Agent / 内容生产 Experiment 实验

Hermes PPT 流水线

Slide-making split into a reviewable, reworkable multi-agent flow. 把做 PPT 拆成可验收、可返工的多 Agent 流程。

  • Intake, spec, research, design, review, delivery as tasks 需求、规格、调研、设计、评审、交付拆成任务
  • Quality gate at every step, rework on failure 每步过质量门,不过关就返工
  • The whole run is recorded on video 全程录成了一条视频
云原生 / 可观测性 Open source 开源

可观测性底座

Metrics, logs, and traces in one stack, feeding agents real data. 把指标、日志、链路串成一套,给 Agent 提供真实数据。

  • One Helm install for collection, storage, query, dashboards Helm 一键部署采集、存储、查询、可视化
  • Metrics, logging, and tracing connected Metrics / Logging / Tracing 打通
  • The data foundation for the AIOps agent 是 AIOps Agent 的数据底座

How I work 我怎么做事

01
例:Hermes PPT 质量门控

Write the acceptance check first 先写验收标准,再动手

Define "done" before building. In the PPT pipeline every stage has a review gate; speed comes from generation, certainty from the gates. 动手前先写清楚"怎样算做完"。做 PPT 的流水线里,每一步都有评审关,不过关就返工,速度交给生成器,靠评审关保证质量。

02
例:MedBridge 报价与确认

AI drafts, a human decides AI 起草,人拍板

The agent quotes on the spot, but price and time lock only after a coordinator confirms; it gives no medical advice and sends emergencies to 120. Agent 当场给出逐项报价,但价格和时间要协调员确认后才锁定;它不给医疗建议,急症直接指引打 120。

03
例:AIOps 诊断报告

Every conclusion traces to evidence 结论要能追到证据

Agent judgments must point back to logs, metrics, or a page number. Each line of the incident report carries the evidence it used. Agent 的判断要能追到日志、指标或原文页码。故障诊断报告里每条结论都挂着它用到的那条证据。

04
例:無門 wumen CLI

One capability, many entry points 一份能力,多个入口

The same logic ships as web UI, MCP server, and CLI, usable by people and callable by other agents. 同一套逻辑同时做成网页、MCP 和命令行,人能用,别的 Agent 也能直接调用。

Articles

Written down 最近写下来的

Writing is where experiments become reusable methods. 方法都是试出来的。把过程和踩过的坑写下来。

AI Agents ai-agent

Creating a High-Quality PPT Video with Hermes and PPT Skill 用 Hermes 和 PPT Skill 制作一条高质量 PPT 视频

A real workflow note on using Hermes AI Agent, profiles, skills, and orchestration to produce a PPT-based video, with the full recording available on Bilibili. 记录我如何用 Hermes AI Agent、profile、skill 和工作流来完成 PPT 内容生产,并把过程录制成一条可复用的视频案例。

Jun 16, 2026 2026/06/16
AI Agents ai-agent

AIOps Agent Architecture Design AIOps 智能运维 Agent 总体架构设计

A structured overview of an AIOps Agent, covering requirements, system architecture, streaming output, context management, automated remediation, and observability. 梳理智能运维助手 3.0 的需求分层、总体架构、技术路线、流式输出、上下文管理、自动修复和可观测性设计。

Jun 5, 2026 2026/06/05
AI Agents ai-agent

LangGraph and MCP Orchestration for an AIOps Agent AIOps 智能运维 Agent 的 LangGraph 与 MCP 编排设计

A deep dive into API design, configuration, LangGraph state machines, federated multi-cluster Agents, scheduling, and MCP tool decoupling. 拆解智能运维助手的接口层、配置层、LangGraph 状态机、多集群联邦 Agent、调度引擎和 MCP 工具解耦方案。

Jun 5, 2026 2026/06/05
All Posts 全部文章 Search 搜索文章

Let's build something in the AI era. 在 AI 时代,一起做点事情