AI Engineering — 10-Lesson Course

Bilingual (English / 中文) course on building applications with foundation models, based on Chip Huyen's AI Engineering: Building Applications with Foundation Models (O'Reilly, 2025).

10 lessons, 2 hours each (1 hour concept + 1 hour hands-on).

Two tracks

Track	Material	Who it's for
Concept	AI_Engineering_10_Lesson_Course_EN.md / AI工程化十节课教程_中文版.md	anyone, no coding — web tools only (ChatGPT / Poe / AI Studio)
Code	the .ipynb notebooks	some technical comfort required — see below

The tracks complement each other; neither is a translation of the other. Concept track builds intuition, code track makes the same ideas reproducible and editable.

How much programming do you need?

Concept track: none. Everything runs in a browser with free tools.

Code track: you don't have to write Python, but you can't avoid touching it. To get through the notebooks you need to be comfortable with:

• running commands in a terminal (the Quick start block below)
• getting an API key from a provider dashboard, or installing Ollama
• opening Jupyter and running cells in order
• editing values inside a Python dict — e.g. changing PROVIDER, or filling in your own scores in Lesson 4
• reading an error message and matching it against the Troubleshooting table

No lesson asks you to write a function from scratch, and no prior Python is assumed. But "no programming background at all" is optimistic for this track — a complete beginner should either pair with someone technical for setup, or do the concept track first and come back.

If the terminal is the blocker, an instructor can set the environment up once and hand out a ready-to-run machine or a hosted notebook. From there the course is mostly run the cell, read the output, change one value, run it again.

Quick start

# 1. Environment (already created in .venv/ — recreate with these two lines if needed)
uv venv .venv --python 3.13
uv pip install --python .venv/bin/python -r requirements.txt

# 2. Register the Jupyter kernel
.venv/bin/python -m ipykernel install --user --name aie --display-name "AIE Course (.venv)"

# 3. Add your key
cat > .env <<'KEY'
OPENAI_API_KEY=sk-your-key-here
KEY

# 4. Launch
source .venv/bin/activate
jupyter lab

In VS Code or an existing Jupyter, just pick the kernel AIE Course (.venv).

Start with 00_Environment_Setup.ipynb — it tests connectivity to every provider and tells you which ones your keys can reach.

Choosing a provider

Every lesson notebook opens with the same config cell. Change one line to switch:

PROVIDER = 'openai'   # 'openai' / 'deepseek' / 'openrouter' / 'ollama'

PROVIDER	Key in .env	Small / big model	Embeddings (Lesson 6)
openai	OPENAI_API_KEY	gpt-5.6-luna / gpt-5.6-terra	✅ text-embedding-3-small
deepseek	DEEPSEEK_API_KEY	deepseek-v4-flash / deepseek-v4-pro	❌ falls back to a local bag-of-words vector
openrouter	OPENROUTER_API_KEY	openai/gpt-5.6-luna / openai/gpt-5.6-terra	❌ same fallback
ollama	none (local, free)	gemma4:e2b-mlx — or any model you pull	✅ nomic-embed-text (pull it first)

All four speak the OpenAI API format, so no other code changes. The rest of each notebook only uses MODEL, MODEL_BIG, and EMBEDDING_MODEL.

Running all ten notebooks costs roughly $0.10 on OpenAI pricing. DeepSeek is cheaper; Ollama is free.

Running fully offline with Ollama

ollama serve
ollama pull gemma4:e2b-mlx      # ~6.5GB, MLX build for Apple Silicon
ollama pull nomic-embed-text    # ~275MB, only needed for Lesson 6

Then set PROVIDER = 'ollama'.

gemma4:e2b-mlx is only the proof of concept — use any local model your machine can handle. It is a small 2B model chosen to show that the notebooks run end-to-end with zero API spend. It does run every cell, but answer quality is well below the cloud models, and it shows most in the evaluation and judging lessons. A bigger local model makes those lessons noticeably better.

To swap it, pull whatever you like from ollama.com/library and put that name in the ollama block of the config cell:

'ollama': {
    'model': 'your-model-here',        # e.g. a 7B/8B chat model
    'model_big': 'your-bigger-model',  # can be the same name if you only have one
    ...
}

Rough sizing — a 4-bit quantized model needs about 1GB of RAM/VRAM per billion parameters, plus a couple of GB of headroom:

Your machine	Practical model size
8GB RAM	2–3B
16GB RAM	7–8B
32GB RAM	13–14B, or a small MoE
64GB+ RAM / dedicated GPU	30B+ — closest to cloud quality

Setting model and model_big to two different local models also makes Lesson 4 (model comparison) and Lesson 9 (small vs. big latency) meaningful again; with one model both sides of those comparisons are identical.

On Apple Silicon, -mlx builds run through Apple's MLX framework and are faster than the generic GGUF builds. On other hardware just drop the suffix.

Lessons

#	English notebook	中文 notebook
1	Introduction to AI Engineering	走进AI工程化
2	Understanding Foundation Models	揭开基础模型的面纱
3	AI Evaluation Fundamentals	AI评估入门
4	AI System Evaluation in Practice	AI系统评估实践
5	Prompt Engineering	提示工程
6	RAG and AI Agents	RAG与AI智能体
7	Model Fine-Tuning	模型微调
8	Data Is King	数据为王
9	Inference Optimization	推理优化
10	AI Architecture and Continuous Evolution	AI架构与持续进化

Repo layout

00_Environment_Setup.ipynb              provider connectivity check — run this first
EN_Lesson_XX_*.ipynb                    code track, English
Lesson_XX_*.ipynb                       code track, 中文
AI_Engineering_10_Lesson_Course_EN.md   concept track, English
AI工程化十节课教程_中文版.md               concept track, 中文
assets/                                 images used by this README
requirements.txt                        shared dependencies
.env                                    your API keys (git-ignored, never commit)

Troubleshooting

Symptom	Fix
no API key found for <provider>	.env missing, or the key name doesn't match the table above
Lesson 6 prints a bag-of-words warning	provider has no embeddings endpoint — switch to openai, or ollama + nomic-embed-text
model "..." not found, try pulling it first	ollama pull <model>
Kernel has no openai module	wrong kernel selected — pick AIE Course (.venv)

Notes for instructors

• Keys live in .env only; .gitignore already excludes it. Never commit a key.
• Model IDs and prices drift. Verify against the provider's pricing page before teaching Lesson 9, which quotes per-million-token rates.