Professional Certificate in How AI Works: Models, Data & Systems

What you will be able to do

• Explain tokens, context windows and why models behave as they do
• Describe training, fine-tuning and alignment and what each changes
• Design a retrieval approach over your own documents
• Choose between prompting, retrieval and fine-tuning with real reasoning
• Build an evaluation plan that proves whether a system works

What is inside

6 modules, 30 lessons. Each module ends in a graded quiz and most carry an assignment.

1. 01

   Inside a Model: Tokens, Parameters and Prediction

   A genuine working understanding of what happens inside a language model, with no mathematics and no coding required. Covers what a token actually is and why tokenisation explains a whole family of surprising failures, what parameters and weights are and what training does and does not put into them, the transformer idea explained conceptually with attention as the mechanism for deciding what matters in context, why the same input can produce different outputs, how temperature and sampling settings really work and what they cannot fix, and what genuinely determines whether a model is good at a particular task. Taught through Halcyon Labs, a fictional company building an internal artificial intelligence assistant over its own product documentation and six years of support history.

   5 lessons · 15 quiz questions

2. 02

   Training, Fine-Tuning and Alignment

   The manufacturing process behind a usable assistant, taught so that you can work out which stage causes which behaviour and therefore which layer to fix. Covers the four stages between a pile of raw text and something you can talk to, what pretraining actually does and on what kind of data, instruction tuning and the sharp limits of what fine-tuning can teach, alignment and learning from human preferences at a concepts level including the distortions that optimising against human approval introduces, what a knowledge cutoff means in practice and why a model cannot tell you what it does not know, where refusals come from and why they are wrong in both directions at once, and what the open weights versus proprietary decision genuinely changes. No mathematics and no coding are required. Taught through Halcyon Labs, a fictional company building an internal assistant over its own documentation and six years of support history.

   5 lessons · 15 quiz questions

3. 03

   Context, Memory and Retrieval

   How a language model gets information at the moment it answers, and why that single mechanism explains most of what these systems can and cannot do. Covers the context window as working memory and what falls out of it, the three different kinds of forgetting people confuse, what products build to compensate, embeddings and semantic similarity explained without mathematics, vector search and the indexes that make it fast, retrieval augmented generation end to end and why it is usually the right first answer for company knowledge, chunking and why it quietly decides retrieval quality, and a field guide to the ways retrieval fails in production. Taught through Halcyon Labs, a fictional company building an internal artificial intelligence assistant over its own documentation and support history.

   5 lessons · 15 quiz questions

4. 04

   Building With Models: Retrieval, Fine-Tuning and Agents

   How working systems are actually assembled around a language model, and how to choose between the options without wasting six months. Covers the ladder of prompting, then retrieval, then fine-tuning and what each genuinely costs, how retrieval augmented generation works step by step and where it fails, when fine-tuning helps and when it is the wrong tool entirely, tool use and function calling in plain terms, what an agent actually is and where agents break, error compounding across multi-step workflows, multimodal systems and their specific blind spots, and latency, cost and reliability as design constraints rather than afterthoughts. Taught through Halcyon Labs, a fictional company building an internal artificial intelligence assistant over its own documentation and support history.

   5 lessons · 15 quiz questions

5. 05

   Evaluation: Knowing Whether It Actually Works

   How to find out whether an artificial intelligence system is genuinely working rather than merely impressive in a demo. Covers why demonstrations mislead and what evaluation actually is, building a test set from real cases instead of imagined ones, measuring accuracy, retrieval quality, faithfulness to source and consistency across runs, running human evaluation that produces reliable numbers, regression testing every time the model or the prompt or the documents change, monitoring a live system and reading the signals users actually give you, and making an honest decision about whether something is good enough to put in front of people. Taught through Halcyon Labs, a fictional company building an internal assistant over its own documentation and six years of support history.

   5 lessons · 15 quiz questions

6. 06

   Safety, Governance and the Systems Around the Model

   The engineering and organisational work that turns a promising assistant into something a company can responsibly put in front of its staff. Covers prompt injection and why untrusted input is the central security problem of these systems, data leakage and what actually leaves your organisation, access control over retrieved documents, guardrails and filtering and the points where a human must stay in the loop, logging and auditability, the governance expectations now forming around artificial intelligence at a concepts level, and a realistic end to end architecture for the Halcyon assistant with its expected failure modes named honestly. Taught through Halcyon Labs, a fictional company building an internal assistant over its own documentation and six years of support history. Carries the course capstone.

   5 lessons · 15 quiz questions · assignment