Level 5 · AI / LLM systems
Shipping a GenAI app: CI/CD, versions and evaluation gates
Release a generative AI application on AWS through a pipeline: define its Bedrock and AgentCore resources as code, pin immutable versions of prompts, guardrails, flows and agents, gate every change on an evaluation and adversarial tests before a human approval, ship with a canary and roll back by moving a pointer, and treat model retirement and data-residency targets as releases too.
What you take away
- A GenAI release is a snapshot, not a commit: code, prompt version, guardrail version, flow or agent version, model id and the evaluation dataset it passed, all defined as code and promoted with the same template from account to account.
And 3 more inside the class, in Pro.
What you will practise
- 1 prediction before the explanation
- 4 practice questions
- 5 cards for your reviews
Cards you add come back in your spaced reviews, when you are about to forget them.
How the class is laid out
The class's sections. What is in them is in Pro.
- The problem
- The simple version
- What one release contains
- The pipeline, stage by stage
- Models retire, and some data cannot travel
- Check your understanding
SAA-C03 and AIP-C01 mark the modules that cover objectives of those AWS exams on the design side. The exam also needs hands-on practice in an AWS account and practice exams: from AWS, only the official 20-question set is free. AIP-C01 is offered only in English, Japanese, Korean and Simplified Chinese. Not affiliated with or endorsed by AWS.