Skip to content
Analogy · visual · math · predict · run · compare · explain

Ship ML and LLM systems
without a math degree.

Courses for working software engineers. Every idea starts as something you already know — a SQL join, an API contract, a code review, a flaky test — then a picture, then a small piece of Python you can run on a laptop. No GPU. No Jupyter.

Start week 0 How to run the exercises ☕ Support this course

21 weeks of ML fundamentals · 16 weeks across LangChain/LangGraph/CrewAI · Laptop-friendly, samples ~8k rows

learn-ml / roadmap

00 Python as glue, NumPy as a typed column

07 Classification, regression, ranking

16 The job pipeline — gate, prod dir, tonight's CSV

LC LangChain, LangGraph, CrewAI — optional frameworks

CloudWave data, one company throughout
21Weeks of ML fundamentals
16Weeks of framework tracks
0GPU required (job path)
8kRows sampled on a laptop

Pick your track

  • ML Fundamentals

    21 weeks (0–20). The job is 0–17. Deep learning 18–20 is optional.

    Analogies first. Formulas only as “math, translated.”

    Start week 0 →

  • LangChain

    7 weeks. Prompts, tools, RAG, a golden file that can fail CI.

    Treat the framework as middleware, not magic.

    Start week 1 →

  • LangGraph

    5 weeks. State machines, branches, checkpoints, human-in-the-loop.

    A workflow engine you already understand.

    Start week 1 →

  • CrewAI

    4 weeks. Roles, tickets, crews.

    Staff a team of agents the way you staff a sprint.

    Start week 1 →

How a week is taught

Box Meaning
Think of it like… Everyday or software analogy. Start here.
If you already write software The mapping to APIs, SQL, reviews, CI.
Engineer mental model How this shows up in a codebase.
Watch out The foot-gun of the week.
Before you run this Predict which metric moves, then compare.
Ship / don’t ship A decision rule, not theory.
Exercise A separate page + a starter.py you run in a terminal.

Lessons are markdown. Exercises are ordinary Python. Read on GitHub Pages, clone the repo when you want to type.

CloudWave

The ML course uses one fake SaaS company. Same customers all the way through.

File Grain Rows
subscriptions.csv one customer ~49k
user_events.csv one event 220k
feature_usage.csv one user × feature × day 160k
feedback.json one comment (JSON Lines) 10k
product_catalog.csv one feature/product 300

Laptop mode samples ~8k customers so a week finishes in a few minutes on 8 GB RAM. Billing is clipped at 2024-11-30 (~49k customers).

Where the files live

Schemas, grain, and download links: CloudWave datasets. The same tables, with extra access notes, are in DATASET_GUIDE.md at the repo root. Load them with lib.course_data.find_data_dir / load_customer_360 — do not hunt for a Kaggle zip.

How to run the exercises → Dataset schemas →

Before beginning, read who this course is—and is not—for.