AI Foundations

1. 01

   The maths you actually need

   Beginner

   Linear algebra, calculus and probability — only the parts that recur. Skip nothing here and everything downstream gets easier.

   • Essence of Linear Algebra Video YouTube 3Blue1Brown 3h Free Watch this before any textbook. It builds the geometric intuition that makes matrices stop being arbitrary.
   • MIT 18.06 Linear Algebra Course MIT OpenCourseWare Gilbert Strang 35h Free The canonical course. Do the problem sets — reading alone will not stick.
   • Essence of Calculus Video YouTube 3Blue1Brown 3h Free Enough calculus to understand what a gradient is and why backpropagation works.

   Build this

   Linear regression from scratch, no libraries

   Implement least squares using only NumPy array operations — no scikit-learn. Derive the gradient yourself, write the update loop, and plot the loss curve.

   The point is not the model. It is that every later abstraction is this, repeated.

   Done when: A notebook where the loss curve goes down and you can explain every line.

   California Housing
2. 02

   Python for AI

   Beginner

   The working subset: NumPy, pandas, and enough software practice that your notebooks survive contact with other people.

   • The Python Tutorial Documentation Python Software Foundation 8h Free The official tutorial is better than most paid courses and always current.
   • pandas — Getting started Documentation pandas 4h Free Work through the 10-minute tour, then the cookbook when you hit something real.
3. 03

   Classical machine learning

   Intermediate

   Regression, trees, ensembles and — more importantly — evaluation. Most production ML is still this, not deep learning.

   • scikit-learn User Guide Documentation scikit-learn 15h Free Read the model evaluation and model selection chapters twice. That is where beginners lose.
   • StatQuest — Machine Learning Video YouTube Josh Starmer 10h Free The clearest explanations of bias-variance, cross-validation and ROC curves anywhere.
   • Kaggle Learn — Intro to Machine Learning Interactive Kaggle 3h Free Short, hands-on, and it runs in the browser.

   Build this

   Titanic — a baseline you can defend

   The standard first competition. Do it properly: establish a dumb baseline first, then justify every change against cross-validated score rather than leaderboard position.

   Write down why each feature helped. That write-up is the actual deliverable.

   Done when: A submission plus a short note explaining what moved the score and what did not.

   Titanic — Machine Learning from Disaster

   Build this

   An end-to-end tabular pipeline

   Take any UCI dataset and build the whole thing: cleaning, feature pipeline, model selection, calibration, and a written error analysis of where it fails and for whom.

   Done when: A reproducible pipeline and an honest error analysis.

   UCI Machine Learning Repository
4. 04

   Deep learning

   Intermediate

   Neural networks from first principles, then the two architectures that matter.

   • Neural Networks: Zero to Hero Video YouTube Andrej Karpathy 15h Free Builds backpropagation and then a GPT from nothing, in code you type yourself. If you do one thing on this path, do this.
   • Practical Deep Learning for Coders Course fast.ai Jeremy Howard 20h Free Top-down: you train a working model in lesson one and learn why afterwards.
   • CS231n — Deep Learning for Computer Vision Course Stanford 30h Free The notes are the reason to come here, even if you never watch a lecture.

   Build this

   A digit classifier with hand-written backprop

   Write the forward and backward passes yourself in NumPy, then reimplement in PyTorch and confirm the gradients match.

   Doing both is what turns autograd from magic into a convenience.

   Done when: Two implementations that agree, and a note on where your manual gradients first went wrong.

   MNIST
5. 05

   Transformers and LLMs

   Advanced

   The architecture behind everything current — read the paper, then build with it.

   • Attention Is All You Need Paper arXiv Vaswani et al. 2h Free The transformer paper. Read it after Karpathy's GPT video, not before — it is far easier once you have built one.
   • The Illustrated Transformer Article jalammar.github.io Jay Alammar 45 min Free The diagrams that make attention click. Read alongside the paper.
   • Hugging Face NLP Course Course Hugging Face 15h Free The practical bridge from understanding transformers to shipping with them.
6. 06

   RAG, evaluation and guardrails

   Advanced

   Retrieval, embeddings, and the part almost everyone skips: proving the thing works.

   • Hugging Face Transformers documentation Documentation Hugging Face 10h Free Reference rather than reading — but know the task guides exist.

   Build this

   A RAG app with a real evaluation set

   Build retrieval over a corpus you actually care about. Then do the hard half: write 50 question–answer pairs by hand, measure retrieval precision and answer faithfulness, and record what breaks.

   A RAG demo takes an afternoon. A RAG system you can defend takes the evaluation set — which is exactly why most portfolios do not have one.

   Done when: A working app, a hand-written eval set, and a table of what each retrieval change did to the numbers.

   Your own corpus, or any open document set

Nothing matches that combination in this path yet.