Section C–C · Learning archive学习档案2023 — 2026

How I got here一路走来

Earlier projects that are not portfolio pieces on their own: tutorials, coursework, hackathons and prototypes. Each one taught me a rule I still build by, and each rule points to the work where it shows today.单独拿出来不够格当作品的早期项目：教程、课程、黑客松和原型。每一个都教会了我一条至今仍在用的规则，每条规则都指向今天体现它的作品。

  1. 2023First LLM apps · York MSc · hackathons第一批 LLM 应用 · 约克硕士 · 黑客松
  2. 2024Rebuilding models · team product拆模型 · 团队产品
  3. 2025Polygraf.ai · PII & policyPolygraf.ai · PII 与策略
  4. 2026Production at Jackson VenturesJackson Ventures 生产环境
A012023Self-started自发项目

First LLM apps第一批 LLM 应用

Python · Flask · OpenAI text-davinci-003

Then当时
Content generators that sent one prompt to GPT-3 and printed the answer, wrapped in small Flask apps.把一个提示词发给 GPT-3 再打印结果的内容生成器，外面包一层小型 Flask 应用。
What it taught me学到的
A prompt that works in a demo is not a feature. There was no test set, so I could not tell a better prompt from a luckier one.在演示里能跑的提示词不等于功能。没有测试集，就分不清是提示词更好还是运气更好。
The rule规则
Measure on a held-out set before claiming it works.先在留出集上测量，再说它能用。
Repos仓库: AI-Content-Generator · Agent-AI-Generator-Version2Shows today in今天体现在 Booking Agent →
A022023–26MSc coursework硕士课程

York MSc in AI约克大学 AI 硕士

Python · Jupyter · scikit-learn

Then当时
Applied AI and machine-learning modules at the University of York, worked as Colab notebooks on real datasets such as renewable energy consumption.约克大学的应用 AI 与机器学习课程，在 Colab notebook 里处理可再生能源消耗等真实数据集。
What it taught me学到的
Evaluation design matters more than the model: split the data before you look at it, report per-class results, and say what the numbers do not cover.评估设计比模型更重要：先切分数据再看数据，按类别报告结果，并说明数字没覆盖什么。
The rule规则
Report recall and precision per category and per language, and state the limits next to the numbers.按类别、按语言报告召回率和精确率，并把局限写在数字旁边。
Repos仓库: AAI-York · AIML-YorkShows today in今天体现在 PII Detector →
A032023Hackathons黑客松

Weekend products周末产品

React · Jupyter · ElevenLabs · Luma

Then当时
LightUp, a concept for transparent charity donations; Gen-AI Hack II, voice and 3D generation experiments.LightUp：让慈善捐款更透明的产品概念；Gen-AI Hack II：语音与 3D 生成实验。
What it taught me学到的
A weekend demo proves the idea and hides the edge cases. What the demo does not handle is where the real work starts.周末做出的演示能证明想法，也会藏住边界情况。演示处理不了的地方，才是真正工作的开始。
The rule规则
Write down what it does not handle, next to what it does.把它处理不了的写清楚，放在它能做的旁边。
Repos仓库: Light-Up · Gen-AI-Hack-IIRead the case阅读案例 LightUp →Shows today in今天体现在 Career OPS →
A042023–24Learning by rebuilding照着重建来学

Inside the models拆开模型看

JavaScript · Python · ImageAI · Kaggle

Then当时
Followed Radu Mariescu-Istodor's self-driving car course (sensors and a neural network in plain JavaScript), ran object-detection experiments on top of the ImageAI library, and explored the OTTO session-based recommender challenge.跟着 Radu Mariescu-Istodor 的自动驾驶课程用纯 JavaScript 实现传感器和神经网络；在 ImageAI 库上做目标检测实验；研究 OTTO 基于会话的推荐赛题。
What it taught me学到的
Writing a network by hand made shapes, weights and memory concrete, which is what I check first when picking a model today. Tutorials are for learning; they are credited, not presented as my work.手写一个网络让形状、权重和内存变得具体，这正是我今天选模型时首先检查的东西。教程用来学习，要注明出处，不当成自己的作品。
The rule规则
Know a model's shapes and memory cost before downloading it.下载模型之前，先弄清它的形状和内存成本。
Repos仓库: Self-Driving-Car-Neural-Networks-and-Machine-Learning · Image_AI · otto-recommender-systemShows today in今天体现在 LLM Arch Cards →
A052023–24Team product团队产品

REEA · Django AI web appREEA · Django AI 网站

Python · Django · JavaScript · shell deploy script

Then当时
A Django web app with a dashboard and landing pages for AI chat workflows, built with contract developers and deployed to a server by script.一个面向 AI 助手流程的 Django 网站，含后台和落地页，与外包开发者合作完成，用脚本部署到服务器。
What it taught me学到的
With several people pushing to one server, the deploy script was the only process. Nothing stopped a broken change, and nobody could say what was live.几个人往同一台服务器推代码时，部署脚本就是唯一的流程。没有东西能拦住坏改动，也没人说得清线上是什么版本。
The rule规则
Tests and a release gate before production; every deploy traceable to a commit.上生产之前要有测试和发布闸门；每次部署都能追溯到提交。
Repos仓库: assistant_aiRead the case阅读案例 REEA →Shows today in今天体现在 IPOT Restaurant Platform →
A062025AI-agent prototypesAI Agent 原型

Generated apps生成出来的应用

TypeScript · React · Drizzle · Replit agent

Then当时
An AI companion chat app and a messaging app, built in days with Replit's coding agent.用 Replit 的编程 Agent 几天内做出的 AI 陪伴聊天应用和消息应用。
What it taught me学到的
Generated code arrives fast and without tests, so no one can change it safely a week later. The speed is real; the missing tests are the cost.生成的代码来得快，但没有测试，一周后谁也不敢改。速度是真的，缺失的测试就是代价。
The rule规则
Write the failing test first, even when an agent writes the code.即使代码由 Agent 来写，也先写失败的测试。
Repos仓库: HR_AI · MessagesShows today in今天体现在 AI Governance Platform →