從入門到上線From first steps to production

AI 工程實戰：把 LLM 系統做到能上線 AI engineering: getting LLM systems to production

這門課分三部分。入門篇用白話和互動示範，搞懂 AI 應用怎麼運作。中級篇從頭做一個真的能用的員工手冊機器人，每章加一個功能。進階篇是前線部署工程師（FDE）把系統放進客戶正式環境時用的方法。每一章都附可以在本機執行的 Python 程式。

The course has three parts. Basics explains how AI applications work, in plain language with interactive demos. Intermediate builds a working employee-handbook bot from scratch, one feature per chapter. The core course covers the methods Forward Deployed Engineers (FDEs) use to put systems into customers' production environments. Every chapter comes with Python code you can run locally.

我是初學者：從入門篇開始I'm new: start with the basics 懂基礎：從中級篇開始Know the basics: start intermediate 我有經驗：從進階篇開始I'm experienced: start the core course 不確定？做自我評估Not sure? Take the self-assessment

我該從哪裡開始Where should I start?

入門篇 · 5 章 · 約 75 分鐘Basics · 5 chapters · ~75 min

會一點程式，沒做過 AIYou code a little, never built with AI

寫過簡單的 Python，但聽到 LLM、API、token、RAG 就頭痛。不需要安裝任何東西，在瀏覽器裡就能動手試。

You've written simple Python, but LLM, API, token and RAG make your head spin. Nothing to install: you can try everything in the browser.

中級篇 · 7 章 · 約 2.5 小時Intermediate · 7 chapters · ~2.5 h

懂概念，想實際做一個You get the concepts, want to build one

知道 LLM、API、prompt 是什麼，想親手做出一個有搜尋、工具、記憶、測試，並且部署成網站的完整小專案。

You know what LLMs, APIs and prompts are, and want to build a complete small project with search, tools, memory and tests, deployed as a website.

進階篇 · 9 章 · 約 4 小時Core course · 9 chapters · ~4 h

已經在寫正式環境的程式You already ship production code

懂 HTTP API、SQL、Git、測試與部署，想知道怎麼把 AI 的 demo 做成客戶願意簽收、維運團隊接得住的系統。

You know HTTP APIs, SQL, Git, testing and deployment, and want to turn AI demos into systems customers sign off on and ops teams can run.

這門課不教 Python 語法、機器學習理論或模型訓練。它假設你用現成的模型 API，把力氣花在模型周圍的工程上。遇到看不懂的名詞，可以查詞彙表。

The course doesn't teach Python syntax, ML theory or model training. It assumes you call hosted model APIs and spend your effort on the engineering around them. Unfamiliar term? Check the glossary.

課程地圖Course map

三個篇章依序銜接：入門篇講概念，中級篇把概念做成一個專案，進階篇處理正式上線的問題。

The three parts connect in order: basics explains the concepts, intermediate turns them into a project, and the core course handles production.

入門篇Basics

中級篇：員工手冊機器人Intermediate: the handbook bot

進階篇Core course

學完之後：做一個自己的作品After the course: build something of your own

讀完課程，最好的下一步是做一個能放進作品集的專案。作品集專案頁面提供六個分級題目、完成標準、README 範本與作品檢查表。想考證照的話，認證與學習路線整理了各家 AI 證照（依官方資訊）與一份 12 週學習計畫。

The best next step after the course is a project for your portfolio. The portfolio projects page offers six graded ideas, completion criteria, a README template and a checklist. If you're considering certification, certifications and learning path compares AI certifications (from official sources) and lays out a 12-week plan.

怎麼使用這門課How to use this course

每章開頭有學習目標，結尾有小測驗。測驗答錯會顯示說明，不計分。
入門篇每一章、中級篇第 2 到 5 章，以及進階篇第 4 到 7 章，都有可以在瀏覽器裡操作的示範。示範用的是和 Python 練習相同的演算法，不是寫死的動畫。
程式碼都在 labs/ 資料夾，完全離線可執行，不需要 API key。入門篇的程式在 labs/beginner/，中級篇在 labs/intermediate/。安裝方式見第 9 章。
右上角可以切換中英文與深淺色。閱讀進度存在你的瀏覽器裡。

Each chapter opens with learning objectives and ends with a short quiz. Wrong answers show an explanation; nothing is scored.
Every basics chapter, intermediate chapters 2–5 and core chapters 4–7 include in-browser demos that run the same algorithms as the Python labs, not canned animations.
All code lives in labs/ and runs offline with no API key. Basics code is in labs/beginner/, intermediate in labs/intermediate/. Setup is in chapter 9.
Switch language and theme from the top right. Your reading progress is stored in your browser.

關於數字About numbers

這門課刻意不提供「準確率提升 X%」「延遲低於 Y 毫秒」這類通用數字。這些數字取決於你的資料、模型與流量，拿別人的數字當目標是常見的錯誤。課程教的是怎麼在你的系統上量出自己的數字（第 7 章）。

This course deliberately avoids generic claims like "accuracy improves by X%" or "latency under Y ms". Those numbers depend on your data, model and traffic, and borrowing someone else's figures as targets is a common mistake. What the course teaches is how to measure your own numbers on your own system (chapter 7).

自我評估：從哪裡開始讀Self-assessment: where to start

每題選最接近你現在實際在做的描述。結果只用來建議閱讀順序，不會上傳。

For each question pick the option closest to what you actually do today. The result only suggests a reading order and stays in your browser.

用 AI 寫程式Coding with AI 偶爾在聊天介面問問題，再貼回編輯器。I occasionally ask a chat assistant and paste the answer back. 每天用補全或 agent 改程式，但很少寫規格或設定專案規則。I use completion or an agent daily, but rarely write specs or project rules for it. 我會先寫規格與測試，讓 agent 做小步修改，並逐一審查 diff。I write specs and tests first, have the agent work in small steps, and review every diff.

處理模型輸出Handling model output 在 prompt 裡要求「請回傳 JSON」，然後直接 json.loads。I ask for "JSON only" in the prompt and call json.loads. 用 schema 或 structured output，但沒有驗證業務規則。I use a schema or structured output, but don't validate business rules. schema 加業務規則驗證，失敗時把錯誤回饋給模型重試，並記錄每次嘗試。Schema plus business-rule validation, with error feedback retries that are logged.

檢索 (RAG)Retrieval (RAG) 還沒做過，或只照教學接過向量資料庫。Never built one, or only followed a vector-DB tutorial. 做過向量檢索，但沒處理精確代號、權限，也沒量過 recall。Built vector search, but not exact IDs, permissions, or measured recall. 混合檢索、權限過濾，並用測試集量 recall@k。Hybrid search with permission filtering, with recall@k measured on a test set.

Agent 與工具Agents and tools 還沒讓模型呼叫過會改資料的工具。I haven't let a model call a tool that changes data. 做過工具呼叫，但沒有狀態保存或人工核准。I've built tool calling, but without saved state or human approval. 有 checkpoint、中斷/恢復、最小權限，以及核准紀錄。Checkpoints, interrupt/resume, least privilege and an approval audit trail.

評測Evaluation 改完 prompt 手動試幾個問題，看起來可以就上。After a prompt change I try a few questions by hand and ship if it looks fine. 有一份測試問題，但不是每次改動都跑，也沒有 baseline。I have a list of test questions, but don't run it on every change or keep a baseline. 測試集來自真實流量，CI 每次都跑並和 baseline 比較。Test cases come from real traffic; CI runs them on every change against a baseline.

客戶與上線Customers and production 習慣拿到寫好的需求單再開始做。I prefer to start from a written ticket. 會和需求方討論，但很少先定義成功指標或上線條件。I talk with stakeholders, but rarely define success metrics or launch criteria first. 會先和客戶定出可量測的成功條件，並規劃監控、交接與回滾。I agree on measurable success criteria first and plan monitoring, handover and rollback.