Artificial Intelligence 2026/27

Materials for two modules of the specialisation course in Artificial Intelligence and Big Data, taught together as one path:

• SAA · Machine Learning Systems: how models learn from data, how they are trained, and how to tell whether they really work.
• PIA · AI Programming: building software with AI: Python and its tooling, language models, retrieval, agents, deployment and automation.

Spanish translation of this repository (maintained by NainIglesias)

Start here

1. Working plan: how to work through the year, the three tiers of each topic, the projects you fork and build on, and how to practise.
2. How you are assessed: the exams, how your grade is calculated, and how contributions count.
3. Improving the materials: report errors, ask for clearer explanations, or fix things yourself.

Roadmap

The year in fifteen stages, in the order we work through them; it adapts as the year goes. Each stage belongs to one module: if you take only one, follow its rows. Two stages build on the other module: building with language models (5, PIA) builds on the note on language models of stage 2 (SAA), and RAG (11, PIA) needs how language models work (10, SAA).

Each topic's README gives its path, with the core items marked ★. The worked examples, and the software you build and run, are repositories of their own. Items marked (coming) are being written, and are linked when they are ready.

Term 1

#	Stage	Module	Topic	Worked examples and deep dives
1	Getting started	PIA	Setting up: Python, uv or conda, Git and GitHub, working with an AI coding assistant	First task: numpy-mnist, forked in week 1. Deep dives: a specs-driven Copilot workflow in VS Code, Copilot customizations
2	What AI and machine learning are	SAA	The map of the year: what AI and machine learning are, and each family of models seen from the outside	Deep dives: classical AI, intelligence without learning, probability and Bayes' theorem
3	Python for AI	PIA	Why Python rules AI although it is slow; program structure, with tic-tac-toe, recommended	Deep dive: compare-python-c-efficiency, to measure it yourself
4	Working with data	PIA	Python data essentials: NumPy and pandas to load, inspect, reshape and combine data, find the problems in a table, and get data from web APIs	Exercises: numpy-mnist
5	Building with language models	PIA	Calling models locally and through an API, prompts, structured output, testing and measuring a component that uses a model (coming). Builds on the stage-2 note Language models from the outside	Worked example: a chatbot with guardrails, on a local or a remote model
6	Supervised learning	SAA	Regression and classification, and how to tell whether a model really works (coming). Builds on the stage-2 note The life of a machine learning project	Worked examples: California housing, King County house sales, thyroid disease. Deep dive: data preparation, from exploration to encoding and scaling (coming)
7	Serving models	PIA	A model behind a web API and a demo, in a container, deployed (coming)	Worked examples: an image classifier served with FastAPI, and the same classifier as a Gradio app

Term 2

#	Stage	Module	Topic	Worked examples and deep dives
8	Neural networks	SAA	How a network learns, networks for tables, convolutional networks, reusing a pretrained network (coming). Builds on the stage-2 note Neural networks from the outside
9	Agents and MCP	PIA	Tool calling, agents that use tools safely, and MCP, a standard protocol to connect models to tools (coming)
10	How language models work	SAA	From words as counts to embeddings, attention and transformers (coming). Builds on the stage-2 note Language models from the outside
11	Retrieval-augmented generation (RAG)	PIA	Grounding a model's answers in your own documents, with answers that cite their sources (coming)
12	Automation and business	PIA	Making AI systems run on their own, and deciding whether they are worth it (coming)
13	Unsupervised learning	SAA	Clustering, dimensionality reduction and anomaly detection (coming)
14	Vision and voice with pretrained models	PIA	Computer vision with OpenCV, YOLO object detection, CLIP embeddings and a face-authentication app. Voice with pretrained models: speech to text and text to speech (coming)
15	Reinforcement learning	SAA	Learning from rewards: agents, environments and Q-learning	Worked example: tic-tac-toe with Q-learning, from the simplest version to an agent that never loses

The February exam covers term 1, and the May exam the whole year: see how you are assessed.