Data & AI
Data Science
Data scientists move from describing what happened to estimating what will happen and what would happen if something changed. The mathematics matters, but the judgement about which question is worth modelling matters more.
Start here
If you cannot yet write SQL and clean a dataset, start with Data Analysis. Everything here assumes you can already get to a clean table, and most people who skip that step stall at Stage 3.
Before you begin
- Comfortable with SQL
- Python and pandas
- Secondary school mathematics
Mathematical foundations
Enough to understand what a model is doing. Not a mathematics degree, and not nothing either.
RequiredStatistics and probability
The actual foundation of the field. People who skip it can call a library but cannot tell when its output is nonsense.
Statistics and probability
The actual foundation of the field. People who skip it can call a library but cannot tell when its output is nonsense.
What to learn
- Distributions
- Sampling and the central limit theorem
- Hypothesis testing
- Confidence intervals
- Bayesian thinking in outline
- Common fallacies
Practice
Next: The mathematics underneath the models.
RecommendedLinear algebra and calculus, in outline
You do not need to derive backpropagation. You do need to know what a gradient and a matrix multiplication are, or the models stay opaque.
Linear algebra and calculus, in outline
You do not need to derive backpropagation. You do need to know what a gradient and a matrix multiplication are, or the models stay opaque.
What to learn
- Vectors and matrices
- Dot products
- Derivatives and gradients
- What optimisation is doing
Next: Working with data in code.
Working with data
The practical layer. Most of a data scientist's day is spent here.
RequiredPython for data science
The lingua franca of the field, and the reason almost every tutorial assumes it.
Python for data science
The lingua franca of the field, and the reason almost every tutorial assumes it.
What to learn
- pandas and NumPy
- Notebooks and their reproducibility problems
- Visualisation
- Reading data from databases and APIs
- Writing code you can re-run in six months
Tools
- Python
- pandas
- NumPy
- Jupyter
- matplotlib
Next: Exploration.
RequiredExploratory data analysis
The stage that determines whether the modelling that follows is worth doing at all.
Exploratory data analysis
The stage that determines whether the modelling that follows is worth doing at all.
What to learn
- Understanding distributions before modelling
- Missing data patterns
- Feature relationships
- Detecting leakage early
- Knowing when the data cannot answer the question
Practice
Project
beginnerAn exploratory analysis
Take a public dataset, explore it thoroughly, and write up what it can and cannot tell you, including the data quality problems you found.
- Python
- pandas
- A visualisation library
You can assess whether a dataset supports a question before committing weeks to modelling it.
Next: Modelling.
Modelling
Start with the simple models. They win far more often than people expect.
RequiredSupervised learning
Regression and tree ensembles solve the large majority of real business problems. Reach for them first.
Supervised learning
Regression and tree ensembles solve the large majority of real business problems. Reach for them first.
What to learn
- Linear and logistic regression
- Decision trees and random forests
- Gradient boosting
- Feature engineering
- Regularisation
- Interpreting coefficients and importances
Tools
- scikit-learn
- XGBoost or LightGBM
Next: Evaluation, which matters more than the model.
RequiredEvaluation and validation
The single most common serious mistake in data science is a model that looks excellent because it was evaluated wrongly.
Evaluation and validation
The single most common serious mistake in data science is a model that looks excellent because it was evaluated wrongly.
What to learn
- Train, validation and test splits
- Cross-validation
- Choosing a metric that matches the decision
- Class imbalance
- Data leakage, the classic silent failure
- Baselines
Practice
Next: Unsupervised methods.
RecommendedUnsupervised learning
Useful for segmentation and exploration, and easy to over-interpret.
Unsupervised learning
Useful for segmentation and exploration, and easy to over-interpret.
What to learn
- Clustering
- Dimensionality reduction
- Anomaly detection
- Validating clusters that have no ground truth
Next: Causal questions.
Experiments and causality
The area that most distinguishes a data scientist from someone who can fit a model, and the area most often skipped.
RequiredExperiment design
Most business questions are causal — 'should we do this?' — and prediction alone cannot answer them.
Experiment design
Most business questions are causal — 'should we do this?' — and prediction alone cannot answer them.
What to learn
- A/B test design
- Sample size and power
- Randomisation and its failures
- Peeking and multiple comparisons
- Interpreting a null result honestly
Practice
Next: Causal inference where experiments are impossible.
AdvancedCausal inference
Often you cannot run the experiment. There are principled methods for that, and guessing is not one of them.
Causal inference
Often you cannot run the experiment. There are principled methods for that, and guessing is not one of them.
What to learn
- Confounding
- Difference in differences
- Instrumental variables in outline
- Propensity scoring
- Stating assumptions explicitly
Project
intermediateAn experiment analysis
Design an experiment for a real decision, compute the required sample size, analyse simulated or real results, and write a recommendation that states its assumptions and limitations.
- Python
- SciPy or statsmodels
You can design an experiment that answers the question that was actually asked.
Next: Delivering the work.
Delivering models and findings
A model in a notebook has changed nothing.
RequiredFrom notebook to production
The gap where most data science value is lost. Bridging it is what makes you employable rather than academically interesting.
From notebook to production
The gap where most data science value is lost. Bridging it is what makes you employable rather than academically interesting.
What to learn
- Refactoring notebooks into modules
- Reproducible environments
- Model serialisation
- Handing over to engineering
- Monitoring a model in production
Tools
- Git
- Docker
- MLflow
Next: Communicating it.
RequiredCommunicating uncertainty
Stakeholders want a number. Your job is to give them one and be honest about how much to trust it.
Communicating uncertainty
Stakeholders want a number. Your job is to give them one and be honest about how much to trust it.
What to learn
- Explaining a model without jargon
- Quantifying uncertainty usefully
- Saying what would change your conclusion
- Recommending an action
- Refusing a question the data cannot answer
Project
advancedA model that informs a decision
Build a predictive model for a real question, validate it properly against a baseline, deploy it somewhere it can be queried, and write a report covering its performance, its limits and its failure modes.
- Python
- scikit-learn
- A deployment platform
You have a model someone could act on, and you can state clearly when they should not.
Where this leads
You do not have to pick one now. These are the directions this path opens up once you are working.
Machine Learning Engineering
Focus on building and serving models at scale.
View roadmapData Engineering
Own the pipelines that feed the models.
View roadmapAI Engineering
Work with foundation models rather than training your own.
View roadmapProduct Management
Use quantitative judgement to decide what gets built.
View roadmapContinue your journey
The paths closest to this one. Skills overlap more than the job titles suggest.
Data Analysis
Turn messy data into a decision someone actually makes.
ViewMachine Learning Engineering
Build models that run in production, not only in notebooks.
ViewData Engineering
Build the pipelines that make everyone else's data work possible.
ViewAI Engineering
Build reliable systems on top of models you did not train.
ViewYou do not have to do this alone
Our programs are free, taught live, and built around the same progression. Join one and work through it alongside other people.