Skip to main content

Module 4: Full-Stack AI Delivery

Learning objectives​

By the end of this module, you will be able to:

  • Draw the architecture of a retrieval-augmented generation (RAG) feature and name the failure point at each stage.
  • Build the feature end to end - ingestion, retrieval, generation, API, UI - using the delivery loop from Module 1 and the specs from Module 2.
  • Write an evaluation set and a scoring function for an LLM-backed feature, and explain why unit tests alone are not enough.
  • Add an evaluation gate to a CI/CD pipeline so a change that degrades answer quality fails the build.
  • Describe what "monitoring" means for a feature whose output is text.

Module content​

Chapter 1: A RAG feature end to end​

Ingestion, chunking, embedding, retrieval, prompt assembly, generation, citation. A concrete build with the stack decisions made explicit, and the four places it breaks in production.

Chapter 2: CI/CD and evaluation gates​

Tests for deterministic code, evaluations for probabilistic output. How to build an eval set, score it, run it in CI, and decide what "regression" means when the answer is a paragraph.

Quiz​

Six questions.

Why "full-stack" is in the title​

A model call is one line of code. Everything around it - data pipeline, retrieval, prompt management, API, interface, evaluation, monitoring, cost control - is the feature, and it is ordinary software engineering with one unusual component in the middle. Engineers who treat the model call as the whole feature ship demos; engineers who treat it as one component ship products.