Generative AI Engineering

Programme overview

This is not a course on using chatbots. It teaches you to build AI features that a business can trust. You'll design the support assistant for Paystream, a Nigerian mobile wallet, using its help centre, 900 labelled support tickets and graded answers from two assistant versions. You'll learn how models work and why they make confident mistakes, call them safely from Python with keys kept out of code, write prompts as versioned specifications, and validate every output against a schema. You'll compare a small model, a large model and a classic text classifier on the same tickets, build retrieval and measure its hit rate, show how grounding with citations and refusals changes answer quality, and check an LLM judge against human grades before trusting it. Finally, you'll redact personal data, defend against prompt injection, cost the options, and write a launch recommendation the evidence supports. Every number in the lessons comes from running the code; live model calls are optional and shown with the Anthropic SDK.

Curriculum

1. 01How Large Language Models Work15 min

   • How large language models workWhat a large language model actually does (predict the next token), what tokens and context windows are, why models sound confident when they're wrong, and what that means for anyone building with them.
2. 02Calling a Model from Code15 min

   • Calling a model from codeCall an LLM from Python through an API, keep your key safe, control the output with system prompts, temperature and token limits, and estimate what a feature will cost before you build it.
3. 03Prompts as Engineering25 min

   • Prompts as engineeringWrite prompts the way engineers write specifications (role, task, context, rules, examples and output format), keep untrusted input clearly separated, and version and test prompts like code.
4. 04Structured Outputs and Validation25 min

   • Structured outputs and validationGet model outputs your code can use, by asking for JSON, validating every response against a schema, and handling the ones that fail instead of trusting them blindly.
5. 05Evaluating an LLM Classifier25 min

   • Evaluating an LLM classifierMeasure an LLM's classifications against human labels, compare a small model, a large model and a classic machine learning baseline, and find where each one fails.
6. 06Embeddings and Search15 min

   • Embeddings and searchTurn text into vectors, find the most relevant documents for a question, measure retrieval with hit rate, and set a threshold for questions the documents can't answer.
7. 07Retrieval-Augmented Generation20 min

   • Retrieval-augmented generationBuild the prompt that turns retrieved articles into a grounded answer with citations and honest refusals, and measure how much retrieval improves answers over a model answering from memory.
8. 08Evaluating Generated Answers25 min

   • Evaluating generated answersBuild an evaluation for free-text answers from human grades, automatic checks and an LLM judge, and find out how far the automated judge can be trusted before relying on it.
9. 09Safety, Privacy and Cost25 min

   • Safety, privacy and costRemove personal data before text reaches a model, defend against prompt injection, keep people in charge of high-stakes cases, and choose a model by weighing accuracy against cost.
10. 10Final Project20 min

    • Final project: Paystream's support assistantPlan your final project, a support assistant prototype with ticket triage and grounded answers, evaluated honestly and made safe, and set up the evaluation harness first.

Assessment and certificate

You are assessed with a final assessment (pass mark 60%) and a final project. Your certificate is awarded when you meet those requirements. Paying for the programme does not award it on its own.