Learn AI Create the Future — with Pranay
Main site

All sections

The whole curriculum, grouped into 14 areas. Start at the top if you are new. If you came here for one thing, filter to its area or search — everything is on this page or one click from it.

Every lesson here is written by .

14 areas 117 sections 1,162 of 1,167 lessons published 10 new lessons every day

Showing Doing the Work — 9 sections, 79 lessons. Show all 14 areas

Doing the Work

The judgement calls nobody teaches — what to build, what to compare against, and how to know your numbers are honest.

9 sections 79 of 79 lessons published

Scoping an ML Project

Deciding what to build, whether machine learning is even the right answer, and how you will know it worked.

11 published

  1. When rules beat machine learning
  2. Turning a vague request into a prediction task
  3. Choosing the label
  4. Can this even be learned?
  5. Choosing the metric that matches the decision
  6. Designing for a human reviewer
  7. Write down the constraints before choosing a model
  8. Check the data exists before promising the model
  9. The one-page ML project brief
  10. Getting knowledge out of a domain expert
  11. Explaining a model to the person who decides

Baselines and Choosing a Model

Start with something dumb that works, then find out honestly whether anything fancier is actually better.

9 published

  1. The baselines you must beat first
  2. Build the whole pipeline with a fake model
  3. Logistic regression, gradient boosting, or a neural net?
  4. Freeze, fine-tune, or train from scratch?
  5. Buying accuracy with size, and when to stop
  6. A comparison that actually proves something
  7. Reading your own errors
  8. Evaluating by slice, not by average
  9. Deciding a model is good enough

Reading Training Curves

A loss curve is a message from your model about what is going wrong. This section teaches you to read it.

7 published

  1. How to read a loss curve
  2. When validation loss rises but accuracy improves
  3. When your validation curve is too noisy to trust
  4. What changing batch size actually changes
  5. How smoothing and log scales mislead you
  6. Plateaus, sudden drops and double descent
  7. Too small a model, or trained too little?

Data Leakage and Results That Are Too Good

A score that looks amazing is usually a bug, not a breakthrough. Here is how to hunt down the leak before someone else finds it.

11 published

  1. Your model scores 99%: what to suspect first
  2. Target leakage: features that contain the answer
  3. Duplicate rows across your splits
  4. Fitting the scaler before splitting
  5. Using the future to predict the past
  6. The same person in train and test
  7. Leakage that survives cross-validation
  8. Choosing features on the full dataset
  9. Overfitting your own test set
  10. Did the model already see your test set?
  11. Hunting a leak with ablations

Reproducibility and Running Experiments

Getting the same number twice, and keeping track of hundreds of runs without losing your mind.

6 published

  1. Report a range, not a single number
  2. Configs instead of constants scattered everywhere
  3. Making sense of two hundred runs
  4. When you cannot reproduce your own number
  5. Designing an ablation that proves your claim
  6. Is this improvement real?

ML Code That Survives

Turning a working notebook into code that another person — including you in six months — can actually run.

7 published

  1. Laying out an ML project
  2. Getting out of the notebook
  3. Asserting on your data before you train
  4. Refactoring a 600-line training script
  5. Reviewing someone's ML code
  6. Reproducing a failure you only see in production
  7. Building a demo that does not break

Reading and Reimplementing Papers

Turning a PDF full of equations into code that runs, and knowing which papers deserve your weekend.

9 published

  1. Reading an ML paper in three passes
  2. Reading a results table sceptically
  3. Decoding the notation
  4. Turning an equation into tensor code
  5. A staged plan for reimplementing a paper
  6. Why your reimplementation is three points worse
  7. Finding the details the paper left out
  8. Reading a research codebase without drowning
  9. Should you adopt this new method?

AI Projects

Build real things. Every project has full code, a dataset and a deployment guide.

9 published

  1. Spam detector
  2. House price prediction
  3. Handwritten digit recognition
  4. Sentiment analysis
  5. Face recognition
  6. Recommendation system
  7. Chat with your PDF (RAG)
  8. Local LLM assistant
  9. Build an AI agent

ML Interview Preparation

What AI and ML interviews actually ask in each round, with worked answers — not career platitudes.

10 published

  1. The ML interview landscape
  2. ML coding interviews
  3. ML theory questions, with answers
  4. Statistics questions, with answers
  5. Deep learning questions, with answers
  6. LLM and GenAI questions, with answers
  7. ML system design interviews
  8. LLM system design interviews
  9. ML case-study interviews
  10. Take-home assignments that pass

Other ways to use this site

Learning paths

An ordered route through the lessons for one job: ML developer, AI engineer, computer vision engineer.

Projects

Build real things with full code, a dataset and a deployment guide.

AI glossary

Every term you keep seeing, defined in one plain sentence first.

Error database

Paste the error you got. Find out what it means and how to fix it.