Portfolio2026 · Chengdu, China中国成都
I build software and the way it feels: the protocol underneath, the app people touch, and the motion and film that explain it.我写软件，也设计它给人的感觉：底层的协议、人手里的应用，以及把它讲清楚的动效和短片。
林光 · two trunks and a beam of light. Click to replay.林光 · 两根树干，一束光。点击重播。
01Hello你好
Rust and TypeScript services that stay correct when the network does not.网络出问题时，依然算得对的 Rust 和 TypeScript 服务。
Desktop, phone and Mac apps that feel native, down to the glass.让人感觉原生的桌面、手机和 Mac 应用，细到玻璃质感。
A mark built from one geometry, and a film cut on the beat.从一份几何长出来的标志，和一部踩着节拍剪的短片。
02Selected work作品
From a one-evening Mac tool to a system that spans three devices. Each page covers the problem, what I built, how it works and what I learned.从一个晚上就能用上的 Mac 小工具，到横跨三台设备的系统。每个详情页都写了问题、我做了什么、怎么做的，以及我学到了什么。
03Lab试一试
Three small, real behaviours from my projects, rebuilt to run in your browser. Sample data only; nothing leaves this page.从我的项目里挑了三个真实的小行为，重写成能在浏览器里跑的版本。只用示例数据，什么都不会离开这个页面。
Click an upstream to take it down. Requests try the waterfall in order with bounded retries; after three failures a circuit breaker opens and the upstream is skipped until a probe succeeds.点一下上游就能把它弄坏。请求按瀑布顺序尝试，重试次数有上限；连续失败三次，熔断器打开，这个上游会被跳过，直到探测成功。
A replay of what llm-provider-kit’s decoder does. Each protocol streams server-sent events with different names and shapes; the decoder turns every frame into the same WireEvent, so the caller writes one loop.重放 llm-provider-kit 解码器做的事。每种协议都用 SSE 流式输出，但事件名和结构各不相同；解码器把每一帧都变成同一种 WireEvent，调用方只需要写一个循环。
Wire · raw SSE线路 · 原始 SSE
Decoded · WireEvent解码后 · WireEvent
What the user sees用户看到的
The reading style of MD_reader, my Mac app: one quiet column, hairline tables, code as cards. Edit the text and the page renders as you type. Press the theme button, the way T does in the app.这是我做的 Mac 应用 MD_reader 的阅读样式：一栏安静的正文、发丝线表格、卡片式代码块。改动左边的文字，右边会边打边渲染。按主题按钮，就像在应用里按 T。
04Craft手艺
My redesign of the Aogiri phone app. Drag the handle, or focus it and use the arrow keys. A neutral base, one accent colour for one meaning, lighter type, and a rim of light on every glass control: bright on top, fading down the sides.我对 Aogiri 手机端做的一次改版。拖动把手，或者聚焦后用方向键。中性底色，一种强调色只表达一种含义，字更轻，每个玻璃控件都有一圈光边：上沿亮，往两侧渐暗。
The full phone story →手机端的完整故事 →
The ember companion from the phone app. Move your pointer and it leans and looks; tap it and it squishes. It breathes on a 4-second cycle and blinks at random, never on a timer you can see.手机端里那团小小的光。移动指针，它会侧身去看；点它一下，它会缩一缩。它按 4 秒一次呼吸，眨眼的间隔是随机的，不会让你看出规律。
The beam never touches a trunk, so it reads as light, not as the third stroke of an N. One geometry drives every icon, the launch animation, this site’s mark and Lumen’s colours. Click a mark to replay it.光不碰树干，所以读作「光」，而不是 N 的第三笔。同一份几何驱动所有图标、开屏动画、本站的标志和 Lumen 的配色。点击标志可以重播。
05Notes札记
Getting something to work takes a day. Getting it to feel right takes weeks, and almost nobody notices those weeks. They only notice when it’s missing. I’ve come to love that last stretch. It’s where a thing stops being mine and starts being something someone else can trust.
把一个东西做到「能用」，也许一天就够；做到「对了」，要好几个星期，而且几乎没人会注意到这几个星期，只有缺了的时候才会被发现。我慢慢喜欢上了这最后一段路：一个东西就是在这里，从「我的作品」变成了「别人可以放心用的东西」。
It’s tempting to report only good news. But when I started saying plainly what didn’t work yet, people trusted my good news more. I would rather show an honest “not yet” than a confident “done” that isn’t true.
只报喜不报忧是很诱人的。可当我开始坦白地说出「这里还没做好」，别人反而更相信我说的好消息。我宁愿给出一个诚实的「还没好」，也不愿给一个看起来很自信、却不真实的「完成了」。
I have been sure about what people needed many times, and wrong almost as often. What they actually wanted was simpler and more human than my plan. A plan is my guess. What people say is the closest thing I have to the truth.
我很多次笃定地以为自己知道别人需要什么，又几乎同样多次地错了。他们真正想要的，往往比我的计划更简单，也更贴近生活。计划只是我的猜测，别人说出口的话，才是我能拿到的、最接近真相的东西。
Small budgets, slow machines, rules I can’t change. Each limit first felt like a wall. Each one later turned out to be the reason the final idea was cleaner. I’ve stopped asking “what would I do without this limit?” and started asking “what is the best thing I can do inside it?”
预算少、电脑慢、改不了的规则，每一个限制一开始都像一堵墙，后来却都成了最终方案更干净的原因。我不再问「如果没有这个限制我会怎么做」，而是问「在这个限制里，我能做到的最好是什么」。
I used to hesitate before big changes. Now I make sure every step can be undone, and then I move quickly. Being brave isn’t about not being afraid of mistakes. It’s about making mistakes cheap.
以前做大改动之前我会犹豫很久。现在我会先确保每一步都能退回来，然后放心地快走。勇敢不是不怕犯错，而是让犯错变得不那么昂贵。
I build a lot of my work together with AI. It makes me faster, but it doesn’t decide for me. Knowing what is worth making, saying no when something is almost right, and owning the result are still my job. If anything, they matter more now.
我的很多作品是和 AI 一起完成的。它让我更快，却不会替我做决定。判断什么值得做、在「差一点」的时候说不、为最后的结果负责，这些仍然是我的事。甚至可以说，这些比以前更重要了。
06About关于我
I’m a student builder in Chengdu. I work across the whole stack, from Rust services to native apps, and I care most about the details people feel but rarely name: an honest error message, a control that answers your hand, an animation that explains what just happened.我是成都的一名学生开发者。我的工作横跨整个技术栈，从 Rust 服务到原生应用；我最在意的，是那些人能感觉到、却很少说出口的细节：一句诚实的报错、一个跟手的控件、一段把刚才发生了什么讲清楚的动画。
These days I also lead a small team of AI coding agents: I write the brief, set what “done” means, and review every result before it ships. Each project page states exactly what my role was.现在我还带着一个由 AI 编程助手组成的小团队：任务说明由我来写，「做完」的标准由我来定，每个结果上线前都由我检查。每个项目页都写清楚了我具体负责什么。
