AI Engineer Path
0 of 0 studied

A hybrid curriculum for a working software engineer

Build first. Learn the theory when the build breaks.

Seven projects, each one hard enough to force the next layer of fundamentals. The courses listed under every project are the ones to open when you hit the wall, drawn from Stanford, MIT, Berkeley, Karpathy, 3Blue1Brown and the primary papers. Tick a course when you have studied it. Progress is saved in this browser only.

Start with Project 1 Foundations that survive
The learning loop Problem leads to Build, Build to Test, Test to Failure, Failure to Learn, Learn back to Problem, and Fundamentals feed the loop from above. Problem Build Test Failure Learn next problem Fundamentals
The loop runs weekly. Fundamentals enter when a build needs them, not years earlier.

Two kinds of knowledge

Frameworks and model families turn over every couple of years. Mathematics, algorithms and systems do not. Spend the years on the left column and treat the right one as current tooling to pick up in a week when it changes.

Slow-changing, worth years

  • Mathematics: linear algebra, calculus, probability
  • Algorithms and data structures
  • Computer architecture and systems
  • Optimization
  • Information theory
  • Distributed systems
  • Numerical computing
  • Research methodology

Fast-changing, learn on demand

  • LangChain and the agent framework of the year
  • Specific model APIs
  • Current prompting techniques
  • Current serving libraries
  • Particular model families

Cross-cutting foundations to keep open through every project

    Where the projects lead

    Each project moves the job title one step. None of them requires finishing the previous one perfectly.

    1. TodaySoftware Engineer
    2. After projects 1 and 2AI Engineer
    3. After project 3AI Systems Engineer
    4. After project 4Efficient and Edge AI Engineer
    5. After projects 5 and 6Embodied AI and Robotics Engineer
    6. After project 7Technical founder or AI systems architect

    How to use this page

    • Start the project before the courses. Open a course only when the build stalls on something you cannot explain. That is the moment the theory sticks.
    • One primary course per project. The first entry in each list is the backbone. The rest are for the specific wall you hit.
    • Papers are read once, fast. Abstract, figures, method, then back to the code. Keshav's three-pass method in the foundations list is enough.
    • Ship something demoable per project. A trained model, a benchmark table, a robot that moves, a customer conversation. No dark-box work.