AI curriculum

Course outline for AWS Engineering for AI Applications

An AWS AI application needs more than a successful model call. Build a bounded Bedrock client against supplied API fixtures and map retrieval, identity, data, and monitoring responsibilities. Compare Bedrock foundation-model access with SageMaker AI model development/deployment, then test throttling, blocked responses, and approved tool actions without creating paid resources.

Technology: Cloud and Platform Engineering
Duration: 4 hours
Modules: 2 · 2 hours each
Level: Level 2

About AWS Engineering for AI Applications

AWS Engineering for AI Applications Course Objectives

Select Amazon Bedrock or SageMaker AI from workload and ownership requirements.
Implement a Bedrock request/response adapter using supplied Converse API fixtures.
Scope model-invocation, retrieval, and tool-execution identities separately.
Map S3 sources, Knowledge Bases, and authoritative application state.
Inspect Guardrails outcomes without treating content filtering as authorization or a truth guarantee.
Test throttling, failure handling, version-bound action approval, and cost/cleanup records.

Pre-requisites

Familiarity with cloud infrastructure, identity, and networking responsibilities.
Basic command-line, configuration, and application deployment skills.

Before this course

Complete the required courses or bring equivalent experience. Recommended courses add useful depth; they are not enrolment gates.

Required preparation

Cloud Foundations for AI Engineers

Recommended preparation

Observability Foundations for AI Systems

Understands compute, network, storage, and identity responsibilities; telemetry helps with operational checks.

Find your place in the learning map

Lab Setup

Programming runtime, editor, and supplied AWS client scaffold with local Bedrock Converse, Knowledge Bases, Guardrails, and IAM fixtures.
Synthetic prompts/documents, dummy identities, SageMaker AI service profiles, throttling cases, and cost inputs; introductory model/API terms are supplied.
No paid model call, model training, live AWS account, production credentials, or resource provisioning is required.

Detailed Course Outline

Module 1: Model access and AWS service boundaries

Amazon Bedrock: Inspect foundation-model access and Converse request/response fixtures; handle throttling and Guardrails outcomes while checking model, region, and API support rather than assuming universal compatibility.
Networking: Map application-to-Bedrock, retrieval, and tool paths; inspect private-connectivity and cross-region data assumptions using current service profiles rather than assuming inference stays in one location.
IAM: Separate model invocation, source retrieval, and tool execution roles; inspect least-privilege allow/deny fixtures without claiming to demonstrate deployed policy enforcement.
Amazon SageMaker AI and compute: Compare managed model development/deployment with Bedrock access; assign application handlers to Lambda or container/EC2 compute using workload, lifecycle, and capacity requirements.

Practical outcome: Implement a supplied Bedrock Converse client adapter, compare Bedrock with SageMaker AI, and review model/retrieval/tool identities and permitted network/data paths.

Module 2: Ground, monitor, and control AI actions

Storage: Use S3 source and artifact profiles for retrieval inputs, dataset versions, and evaluation evidence; record retention, access, and deletion responsibilities.
Knowledge Bases and data stores: Inspect Bedrock Knowledge Bases retrieval and citation fixtures; distinguish source ingestion and derived retrieval from authoritative application records and approved tool mutations.
Observability: Inspect CloudWatch-oriented request evidence, latency, throttling, token/cost counters, and blocked responses; redact supplied sensitive fields and distinguish execution success from answer correctness.
Automation: Compare Bedrock AgentCore hosting/tool-service roles with application-owned execution; use versioned approvals and bounded retries around mock actions, checking final state, cleanup, and cost assumptions.

Practical outcome: Map S3 sources and Knowledge Bases to retrieval responsibilities, inspect Guardrails and citation fixtures, and define token accounting, bounded retries, and approved tool execution.

Practical exercise

Produce a tested mock Bedrock client, AI service diagram, permission matrix, request evidence, and cost/cleanup plan using supplied version-labeled AWS artifacts.
Accept a valid model response; handle throttling and a Guardrails-blocked response explicitly, and reject an out-of-scope model or source request in policy fixtures.
Inspect a Knowledge Bases retrieval/citation fixture, keep authoritative state separate, and verify changed tool-action parameters invalidate prior approval.
Inject an unavailable-service case; verify bounded retries, redacted monitoring evidence, expected final mock state, and token/storage/capacity cost assumptions.

What you will produce

Test a bounded Bedrock client and design an AWS AI application with explicit retrieval, identity, action, monitoring, and cost responsibilities.

How the work is checked

The selection distinguishes Bedrock model access, SageMaker AI model lifecycle, retrieval services, and authoritative application state.
Client fixtures handle valid, throttled, and blocked responses; policy fixtures reject out-of-scope model/source access.
Changed tool-action parameters fail version-bound approval; Guardrails outcomes are not treated as tool authorization or proof of truth.
Failure records show bounded retries, redacted evidence, expected mock state, and explicit cost/cleanup assumptions.

Four-hour Bedrock-centered client and architecture lab with prepared fixtures. SageMaker AI, Knowledge Bases, Guardrails, and AgentCore receive bounded role comparisons, not full implementations. Check current model/region/API availability and quotas before a live lab. Fixtures do not prove deployed IAM enforcement, answer correctness, or production capacity; no training or paid resources are included.

When another approach is better

Use a simpler hosted API when AWS-specific integration and operating requirements do not justify the extra service setup.
Do not choose SageMaker AI model training/deployment when managed foundation-model access meets the need.
Do not grant model or retrieval identities unrestricted tool mutations, or treat Guardrails as application authorization.

How we train

Ask about exercises, instructor feedback and the prior knowledge you need. Tool-specific courses marked provisional may become modules in a broader course.

Discuss your learning goals