Hi, there! 👋
Engineer building AI agents and system design, writing about work, investing, and better decisions.
I'm smian. I came up through AI agent development, DevOps, microservices, and intelligent routing. Right now I work on AI agent workflows and system design: AIOps, CI/CD, traffic routing, model serving, and the reliability details that decide whether an agent survives real work. My day-to-day stack is Python, Go, JavaScript, Shell, and Vue.
- ✍️ I write on my blog about Agent workflows, engineering practice, career choices, investing, and decision making. I prefer notes that come from something I have built, used, or paid for.
- 📈 Outside code, I spend a lot of time thinking about cash flow, investing, and personal systems.
- 💬 For consulting, add me on WeChat and include the problem you want to solve. A little context upfront makes the conversation much better.
What this site is about
This is my home base for work, writing, and consulting. The engineering side comes from my work on platforms and AI agents development, while the writing reaches a little wider — covering AI agents, work, investing, career planning, daily notes and reflections, and how I make decisions when the answer isn't sitting neatly in a document. It's where I share my thoughts on engineering and personal growth.
Daily AI Signal
AI News
A daily feed for tracking AI tools, agent workflows, model updates, infrastructure shifts, and the original links behind each event.
Browse AI NewsDaily AI News
Subscribe to AI News
Daily AI signal for builders: tools, agents, models, infra, product shifts, and the links behind each event.
No spam. Every issue links back to the original sources.
Experience
QQ Music
AIOps, agents, microservices, DevOps, and traffic routing — turning operational signals into automated diagnosis and self-healing actions.
Tencent Meeting
Built the environment governance system and CI/CD pipelines, making test environments consistent and releases safer and more repeatable.
R&D Engineer · Innovation Research Institute
Worked on intelligent routing algorithms, large language models, model-training acceleration, and alert/anomaly prediction for network systems.
Development Engineer
Built the knowledge base and an AI assistant bot, helping creators answer and serve their users automatically.
Education
M.S. in Computer Technology
Research focused on intelligent routing, large language models, and network traffic prediction.
Projects
对话式 MBTI 人格分析 —— 没有一道选择题，打开就是聊天。心理学里的社会期许偏差（Social Desirability Bias）告诉我们，人填问卷时会无意识地美化自己。talk-mbti 换了个思路：通过自然对话，从你的措辞、第一反应、决策逻辑里提取真实的行为信号，聊完直接生成四维度百分比、认知功能栈、超能力和成长盲区的完整报告。基于 Gemini 2.0 Flash 驱动，Cloudflare AI Gateway 代理，国内无需梯子可用。
Hermes Agent 源码解剖 —— NousResearch/hermes-agent 是一个 33 万行 Python 代码的开源 AI Agent 框架，从 CLI 交互到 11 平台消息网关到 RL 训练数据采集，全部内置在一个单体仓库里。这个项目是对其源码的系统化技术分析，逐模块拆解，写了 8 篇技术文档，配有大量手绘风格技术插图。从 Agent 核心循环、Tool Registry、多 Provider 适配到上下文压缩、消息网关、Memory 与 RL 训练。最后一篇将 Hermes Agent、OpenClaw、Claude Code 三个框架放在一起做了 13 个维度的对比。
Claude Code 源码解剖 —— 2026 年 3 月 Anthropic 意外泄露了 Claude Code 的完整 TypeScript 源码（51.5 万行）。我对这份源码做了系统化的技术分析，逐模块拆解，写了 13 篇技术文档（中英双语）。从架构设计到 Agent 循环的六阶段实现，从三层消息压缩体系到 6,300 行的权限系统，从五层记忆加载到 40+ 工具的执行流水线。同时从 82 个 feature flag 里挖出了 Anthropic 尚未发布的功能蓝图。
Blog
Selected writing on AI tooling, investing, career growth, and engineering practice.
自然语言是新时代的编程语言，编程语言是新时代的汇编
Claude Code 主席 Boris 的播客采访信息量很大。自然语言正在变成新的编程接口，代码越来越像汇编，而软件工程师这个岗位正在向 Builder 演化。
你必须开始输出的七个理由
持续输出不只是为了赚流量的钱。它会强迫你学习、训练 Sale 和 Storytelling、积累数字资产、拓宽信息源，并帮你摆脱单点收入风险。
做事的本质是做人
真正开始带队之后，我越来越相信做事会暴露一个人的内核。主业是训练密度最高的地方，高质量底线会迁移到你做的所有事情上。
对流量和曝光的新认识
曝光不是虚荣，而是让价值被看见的基本能力。无论在公司内部还是外部，你都需要主动设计可见性，让自己的工作变成可积累的资产。
深入分析OpenClaw的爆火
深度解析OpenClaw为什么能在短时间内爆火，从泛流量、垂直流量、大厂跟进时间线等多个维度分析，以及我们能从中学到什么。
Daily US Stock Signal
Stock News
A daily US market feed covering index context, selected stock moves, short-interest signals, IPO calendar notes, and the source data behind each issue.
Browse Stock NewsDaily US Stock News
Subscribe to Stock News
Daily US market notes: index context, selected stock moves, short-interest signals, IPO calendar notes, and source data behind each issue.
No spam. Every issue links back to the original sources.