The Modelwright Project
An open, project-based path to becoming a Machine Learning Engineer.
The Modelwright Project (Modelwright for short) is a free, self-directed curriculum for people who want to learn machine learning seriously, all the way from foundations to research-grade and production-grade work. It is built on a simple belief: you do not learn machine learning by watching it, you learn it by building it. Every concept here is earned by writing code, reimplementing the mechanism from scratch, and shipping a real artifact you could put in front of an employer.
It takes its inspiration from The Odin Project, the open curriculum that has taught hundreds of thousands of people web development. The Odin Project’s insight was never about writing the best tutorials. It was about sequencing the best existing material in the world and surrounding it with projects you cannot fake your way through. Modelwright does the same thing for machine learning, and adds the one ingredient ML needs that web development does not: a relentless emphasis on understanding the internals, because in ML you can finish a course and still not understand how the model actually learns.
Who this is for¶
Modelwright is for anyone who wants to become a machine learning engineer and is willing to do the work. It suits:
Software engineers moving into ML.
Self-taught learners who want depth, not just the ability to call a library.
Students who want a practical complement to formal study.
Working practitioners who want to fill the gaps between “I can use the tools” and “I understand and can rebuild the tools.”
What you need before you start: comfort writing code, ideally in Python, and a willingness to engage with mathematics when a model calls for it. You do not need a mathematics degree. The curriculum teaches the math you need at the moment you need it, and points you to foundational resources if your background is thin (see the math module in Phase 0). If you have never programmed before, learn programming first, then come back.
How it works¶
The curriculum is organized into six phases, moving from foundations to original research-level work. Each phase contains modules, and each module contains units, a unit being roughly one focused session of work.
You move forward by passing gates. A gate is proof that the learning actually landed, and there are three kinds, used in increasing severity:
Concept checks. Short questions you answer without notes. They catch passive watching, the failure mode where a video feels clear but nothing stuck.
Build gates. After you have used a mechanism, you reimplement it from scratch and make it pass tests. You do not advance past attention until you have written attention yourself. You do not advance past gradient descent until your own code trains a network.
Ship gates. You deliver a real, tested, documented artifact: working code, a written report, a deployed service, a reproduced result. Every module has at least one, and every phase ends in a capstone.
Progression is gated, not timed. You advance when the gate is green, however long that takes. The gates are the entire point. Done honestly, this is a multi-year body of work, which is the correct timescale for genuine expertise. You walk it one module at a time.
The four commitments¶
Everything in Modelwright follows four rules:
Experience it, then build it. You always use a technique first, with a library, seeing it work and building intuition, before you implement it from scratch. You earn the right to build something by first understanding what it is for. This is why the very first thing you do is train a real model in a few lines, not assemble a neural network from nothing. Accessible entry, expert ceiling, in that order.
Curate, don’t rewrite. The world already has extraordinary free material: top university courses, the best textbooks, definitive explainer videos, and the original research papers. Modelwright links the best of it with surgical precision, pointing you to exact lectures, chapters, and sections, and tells you what to skip. The original writing here is the connective tissue: why each thing matters, how the pieces fit, and the project specifications.
Understanding is proven by building. After you have used a load-bearing concept, you reimplement it from scratch. This is what separates someone who can use a framework from someone who can debug it, extend it, and reason about it when it breaks. The using comes first; the building is how you prove it really landed.
Everything ships. Projects are not exercises, they are artifacts. By the end of the curriculum, the portfolio you have built is your credential.
Each module follows the same shape¶
So you always know what you are looking at, every module is structured identically: why it matters, what you will be able to do afterward, the curated learning path, knowledge checks, the from-scratch build (where applicable), and the project that proves you can apply it. Optional deeper material and the canonical references close each module.
Where to start¶
New here? Read The Curriculum Map to see the whole journey.
Ready to begin? Start with Phase 0.
Welcome to the workshop. Let’s build.