AWS AI Practitioner

AI is a tool for certain problems, not all of them. You learn when machine learning fits and when a simple rule does better, and meet AWS’s ready-made AI services for vision, language, speech and documents.

• Problems ML solves well, and problems it does not
• Prediction, classification and recommendation examples
• Managed AI services for images, text, speech and documents
• Build, buy or customize
• Costs and risks of an AI project
• Defining a success measure

You build A use case assessment that rates five ideas and picks the best fit for AI.

A model is one step in a longer process. You follow the lifecycle from data collection through training, evaluation, deployment and monitoring, and learn what each step needs from the others.

• Collecting and preparing data
• Features, training and tuning
• Evaluating models with basic metrics
• Deployment and inference options
• Monitoring and retraining
• MLOps and SageMaker at an overview level

You build A lifecycle map for one imagined ML project, with an owner for each step.

Generative AI can write, summarize and create images, but it has hard limits. You see how foundation models are built, what tokens, embeddings and context windows are, and why models sometimes make things up.

• Foundation models and large language models
• Tokens, embeddings and context windows
• Transformers and diffusion models in outline
• What generative AI does well
• Hallucination and other limits
• Multimodal models

You build A short explainer that shows a team what a model can and cannot be trusted to do.

Turning a model into an application involves choices. You compare AWS services for building with generative AI, weigh model selection factors, and learn how retrieval and agents give a model access to your own information and tools.

• Amazon Bedrock and SageMaker JumpStart
• Choosing a model: quality, speed, cost and licensing
• Inference settings and their effect on output
• Retrieval-augmented generation and vector databases
• Agents that use tools
• Cost drivers of a generative AI application

You build A design sketch of a question-answering assistant that names its model, data source and AWS services.

How you ask decides what you get. You write clear prompts, add examples and structure, learn the common techniques, and recognize prompt attacks and the ways to limit them.

• Clear instructions, context and format
• Zero-shot and few-shot prompting
• Step-by-step reasoning prompts
• Prompt templates
• Prompt injection and jailbreaking
• Testing and improving a prompt

You build A prompt template tested on ten inputs, with notes on what improved it.

A general model may not fit your task. You learn how models are pre-trained and fine-tuned, what data that needs, and how to judge output with metrics and with people.

• Pre-training, fine-tuning and continued training
• Instruction tuning and feedback from people
• Preparing data for customization
• Automatic metrics for generated text
• Benchmarks and human evaluation
• Judging whether a business goal is met

You build An evaluation plan that tests a model against a business goal.

Responsible AI is a set of working habits. You learn about fairness, bias, safety, privacy and transparency, why explainable models matter, and tools that help check and document a system.

• Fairness, bias and representative data
• Safety, reliability and veracity
• Explainability and transparency
• Human oversight and review
• Model cards and documentation
• Tools for bias detection and content safety

You build A responsible AI checklist applied to one example system.

AI systems add new places for data to leak and rules to bite. You learn who is responsible for what, how to protect data and access, and how governance and compliance requirements shape an AI project.

• Shared responsibility for AI systems
• Access control and encryption for AI data
• Data lineage, retention and source tracking
• Threats specific to AI systems
• Governance frameworks and regulations
• Auditing and compliance tools

You build A security and governance review for one AI use case.

Combine the whole subject in one document. You pick a use case, decide whether and how to use AI, outline the data, model and risks, and present a plan a non-specialist can follow.

• Choosing and framing a use case
• Selecting an approach and AWS services
• Planning data, evaluation and monitoring
• Covering responsible AI and security
• Presenting the plan

You build A short AI project plan covering approach, data, risks, security and success measures.