我关注如何把真实的问题转化成可运行、可验证的 AI 应用，从问题判断和工作流设计，到原型构建、验证与迭代。
我尤其关注 AI 从 Demo 走向真实使用之前的阶段：问题是否值得解决、AI 应该介入哪里、结果如何验证，以及哪些决定仍然需要由人完成。
我主要通过真实项目学习 AI 应用开发，在构建、踩坑、验证和复盘中持续补充技术基础，积累产品判断与工程经验。
-
真实问题与产品判断
先理解用户、场景和约束，再判断 AI 是否真的适合解决这个问题。 -
AI 应用与工作流原型
把模糊需求拆成能够运行和检查的流程，而不只是做一个聊天界面。 -
Human-in-the-loop
对高风险、不确定或不可逆的操作保留人工确认，让 AI 有明确的行动边界。 -
知识与检索型 AI
实践文档解析、检索、引用和结果验证，并关注回答背后的证据是否可靠。 -
生产前验证
通过测试、失败场景、部署和实际演示，寻找 Demo 与真实使用之间还缺少什么。
| 项目 | 我在解决什么 |
|---|---|
| ArrivalReady | 探索如何用证据而不是单纯的模型判断，检查国际访客接待准备情况，并在关键结论前保留规则检查和人工复核。 |
| OneCase | 将零散的社区居民反馈整理成可编辑、可追溯的事项，并在正式进入后续流程前保留人工确认。 |
| EasyPhone AI | 面向海外华人家庭的手机使用辅助原型，把复杂操作拆成更小、更可逆的步骤，并探索诈骗风险场景下的安全边界。 |
| LOOM | 一个用于实践 AI 工作流、RAG、可视化流程和对话交互的全栈项目。 |
| TeachOps | 探索多 Agent 如何辅助教学设计预审，以及在信息和证据不足时如何阻止流程过早向下执行。 |
更多实验和原型，包括 LifeOPS，可以在我的个人网站中查看。
Observe → Build → Verify → Reflect
Observe
先理解问题、用户和真实约束，而不是从“能用什么模型”开始。
Build
使用 AI 编程工具辅助开发，尽快把想法变成可以运行和交互的版本。
Verify
通过测试、失败场景、部署结果和实际演示检查假设，而不是把“能够运行”等同于“已经解决问题”。
Reflect
记录踩过的坑、错误判断和实现中的不足，再针对暴露出来的问题补知识、调整方案并继续迭代。
我把原型、模拟 Demo 和真实世界验证看作不同程度的证据，也尽量不把尚未验证的能力描述成已经完成的成果。
AI-assisted Development · AI Workflows · RAG · Human-in-the-loop · Git / GitHub
Python · TypeScript / JavaScript · Vue · Next.js
我会使用 Claude、Codex 等 AI 编程工具辅助理解代码、实现功能、调试和迭代，同时尽量理解系统为什么这样工作，而不仅仅让代码能够运行。
正在继续巩固 Python 与计算机基础，并学习 C++。
相比单纯积累更多框架和工具，我更希望通过项目逐步补齐基础，并建立能够迁移到不同问题上的判断能力。
- 🌐 个人网站
- 🧩 案例研究
- 📝 关于 AI 产品与构建的笔记
- 💬 联系我 / 索取简历
English version
I'm learning how to turn real-world problems into working, testable AI applications, from problem framing and workflow design to prototyping, validation, and iteration.
I'm especially interested in the stage between a Demo and real-world use: whether a problem is worth solving, where AI should be involved, how results can be verified, and which decisions should still remain with people.
Most of my learning happens through building real projects. I build, run into problems, verify assumptions, reflect on what failed, and then strengthen the technical foundations that the project exposed.
-
Problem framing and product judgment
Understanding users, scenarios, and constraints before deciding whether AI is actually the right tool. -
AI application and workflow prototypes
Turning ambiguous requirements into workflows that can be run, inspected, and tested. -
Human-in-the-loop systems
Keeping explicit human approval around high-impact, uncertain, or irreversible actions. -
Knowledge and retrieval-based AI
Practicing document processing, retrieval, citations, and result verification. -
Pre-production validation
Using tests, failure cases, deployments, and demos to understand what is still missing between a prototype and real-world use.
| Project | What I'm exploring |
|---|---|
| ArrivalReady | Exploring evidence-based readiness checks for international visitor reception, with rule-based checks and human review around important conclusions. |
| OneCase | Turning fragmented resident feedback into editable and traceable community issues, with human confirmation before formal workflow entry. |
| EasyPhone AI | A phone-assistance prototype for overseas Chinese families that breaks complex tasks into smaller, reversible steps and explores safety boundaries in scam-risk scenarios. |
| LOOM | A full-stack project for practicing AI workflows, RAG, visual flow building, and conversational interaction. |
| TeachOps | Exploring how multi-agent systems can support teaching-design review and prevent downstream work when information or evidence is incomplete. |
More experiments and prototypes, including LifeOPS, are available on my portfolio.
Observe → Build → Verify → Reflect
Observe
Start with the problem, the user, and real constraints rather than starting with a model or framework.
Build
Use AI coding tools to turn ideas into working and interactive prototypes quickly.
Verify
Test assumptions through failure cases, deployment results, and actual demos instead of treating "it runs" as proof that the problem has been solved.
Reflect
Document mistakes, implementation problems, and weak assumptions, then learn what is missing and iterate.
I treat a prototype, a simulated demo, and real-world validation as different levels of evidence, and I try not to present unvalidated capabilities as finished results.
AI-assisted Development · AI Workflows · RAG · Human-in-the-loop · Git / GitHub
Python · TypeScript / JavaScript · Vue · Next.js
I use AI coding tools such as Claude and Codex to help me understand code, implement features, debug, and iterate. At the same time, I try to understand why the system works rather than stopping when the code happens to run.
I'm strengthening my Python and computer science fundamentals while learning C++.
Rather than collecting more frameworks, I want to use projects to strengthen my foundations and gradually build judgment that transfers across different problems.