13 lessons · 5 parts · no cloud GPU required

Build AI models small enough to run on your own device.

Most AI courses teach you to call somebody else's model over the network. This one goes the other way. You will train models from scratch, shrink them until they fit in a phone's memory, and finish with a voice assistant that works with the network switched off entirely.

Start Lesson 01 → See the whole map Set up your machine

the whole course 3.4 GB → 180 MB → in your pocket
13
lessons, each with a project you actually run
~20 h
reading, plus your own training time
0
cloud GPUs required — a laptop is enough for every project
1
offline voice assistant at the end

Who this is for

You already write software. You may already have built things with AI APIs and agents. What you have not done is open the model itself. This course assumes exactly that: strong engineering, zero machine learning.

You should be comfortable with

  • Python: functions, classes, virtual environments.
  • The command line, and reading an error message calmly.
  • School-level maths. Multiplication and a little algebra.

You do not need

  • Any previous machine learning.
  • University mathematics. No proofs appear in this course.
  • An expensive GPU. Everything runs on an ordinary laptop.

How to read the diagrams

Almost every idea in this course is explained twice: once in words, once in an animation you can step through at your own speed. The animations follow one strict rule — colour always means the same thing, in all thirteen lessons. Learn these six colours now and every later diagram will read faster.

data moving forward weights, the things we learn gradients moving backward loss, error, the thing we shrink output, correct, finished frozen or discarded
Using the animations

Each figure has Play, Next and Previous controls, and a caption that changes with the picture. Nothing moves faster than you allow. If your system asks for reduced motion, figures start at their final state instead of playing.

The map

Each part answers one question. Read them in order — later lessons depend on earlier ones.

Final project · Lesson 13

A voice assistant with the network switched off

Everything in this course points at one object: a small speech recogniser, a small language model and a small speech synthesiser, running together on one device, fast enough to feel like a conversation. No API key. No server. Aeroplane mode.

It is a genuinely hard engineering problem, and it is a much more interesting one than wiring three cloud APIs together.

Ears
streaming ASR
Brain
quantized small LM
Voice
streaming TTS
Budget
< 1.5 s voice-to-voice

One honest warning

This field moves quickly and it is not tidy. There is no single agreed course for on-device model engineering, which is exactly why this one exists. Some of what you will read here — especially full-duplex speech in Lesson 13 — sits close to current research rather than settled practice.

Where something is uncertain, this course says so instead of pretending. Where a number is measured, it names the source. And where a technique sounds impressive but rarely helps in practice, it says that too.

Start here

Lesson 01 needs nothing installed. Read it first, then set up your machine with the toolkit page when you reach the project.