AWS · AIF-C01 · STUDY GUIDE

AWS Certified AI Practitioner (AIF-C01) — Complete Study Guide

For business analysts, project managers, developers, and anyone who works with or alongside AI solutions on AWS. Passing AIF-C01 shows you can pick the right AWS service, technique, or safeguard for a described AI, machine learning, or generative AI scenario — no coding or model-building experience required.

Learn → Practice → Apply. Build Smart. Scale Fast.

Expect scenario questions of the form "a company wants X — which AWS service or approach fits?" where the discriminator is usually a keyword: labeled data points to supervised learning, external knowledge without retraining points to RAG, filter harmful content points to Guardrails, and so on. This guide expands concise revision notes into a complete, exam-ready reference with fuller explanations, worked context, and structured summaries.

How to Use This Guide

The guide is organized into five parts that mirror the official exam domains exactly, in official order and weight:

• Part 1 — Fundamentals of AI and ML (20%)core AI/ML concepts, the ML lifecycle, AWS managed AI services, and Amazon SageMaker.
• Part 2 — Fundamentals of Generative AI (24%)generative AI concepts, foundation models, Amazon Bedrock, Amazon Nova, and Amazon Q.
• Part 3 — Applications of Foundation Models (28%)model selection, prompt engineering, inference parameters, RAG, agents, fine-tuning, and evaluation.
• Part 4 — Guidelines for Responsible AI (14%)responsible-AI dimensions, guardrails, interpretability, and generative-AI risks.
• Part 5 — Security, Compliance, and Governance for AI Solutions (14%)shared responsibility, governance, data lineage, and the AWS security service family.

Study Parts 2 and 3 hardest — together Generative AI (24%) and Applications of Foundation Models (28%) are more than half the exam. Work through the parts in order: Part 1 builds the vocabulary the rest of the guide relies on.

Callouts flag material worth extra attention:

• Exam Tippoints that are frequently tested or commonly confused; read these closely before exam day.
• Key Takeawaythe single most important idea from a section, distilled into one or two lines.
• Rememberhandy mnemonics and one-line reminders to lock a concept into memory.
• Notesupporting detail, nuance, or a clarification worth keeping in mind.

Suggested approach: read a section, study its diagram or table, then re-read only the Key Takeaway and Exam Tip callouts the day before your exam for rapid revision. Finish with the Exam-Day Checklist as your final-48-hours review sheet.

Part 1: Fundamentals of AI and ML (20%)

This domain builds the conceptual foundation: what AI, ML, and deep learning are, how models are trained and evaluated, and which AWS managed AI services and SageMaker capabilities implement each step of the ML lifecycle.

Introduction to Artificial Intelligence (AI)

Artificial Intelligence (AI) is a field of computer science dedicated to solving problems that normally require human intelligence — reasoning, perception, learning, and decision-making.

Common AI applications:

• •
  Image creation — generating new images from a description or prompt.
• •
  Image recognition — identifying objects, people, or scenes in an image.
• •
  Speech-to-text — converting spoken audio into written text.
• •
  Learning systems — software that improves its behaviour from experience/data.

How Does AI Work?

At a high level, every AI system follows the same simple pipeline:

Training Dataset → Train Model → AI Model

• •
  Training dataset: provide raw data to the system.
• •
  Train model: the model learns patterns from that data.
• •
  AI model: the trained model makes predictions or decisions on new inputs.

The AI Hierarchy (How the Terms Relate)

These four terms are nested — each is a subset of the one before it. Understanding this relationship is essential for the exam.

Real-World Examples

AI Capabilities and Components

AI is the development of intelligent systems capable of performing tasks that require human intelligence.

The components of an AI solution stack:

• •
  Data layers — collect vast amounts of data.
• •
  ML framework and algorithm logic — engineers work together to understand the data and choose a framework.
• •
  Model layers — implement the model and train it.
• •
  Application layers — how you serve the model and its capabilities to your users.

Machine Learning and Deep Learning

Machine Learning:

• •
  A type of AI for building methods that allow machines to learn.
• •
  Makes predictions based on the data used to train the model.

Deep Learning:

• •
  Uses neurons and synapses (like the human brain) to train a model.
• •
  Processes more complex patterns in data.
• •
  Requires a large amount of input data.
• •
  Requires a GPU for efficient training.

Network structure: an input layer feeds one or more hidden layers, which produce the output layer.

Neural Networks and Transformers

Neural Network:

• •
  Nodes are connected together.
• •
  Nodes are organized in layers.
• •
  As the network sees a lot of data, it identifies patterns and changes the connections between nodes.
• •
  Nodes "talk" to each other by passing data to the next layer.

Transformers Model (LLM):

• •
  Able to process a sentence as a whole instead of word by word.
• •
  Faster and more efficient text processing.
• •
  Gives selective importance to specific words in a sentence (attention).

Common Model Types

Training Data

• •
  To train a model, you must have good data.
• •
  Garbage in ⇒ garbage out.
• •
  It is the most critical step to building a good model.

Two key distinctions: labeled vs. unlabeled data, and structured vs. unstructured data.

Supervised Learning

• •
  Learns a mapping function that can predict the output for new, unseen input data.
• •
  Needs labeled data.
• •
  The model learns from labeled data.

Training vs. Validation vs. Test Set:

• •
  Training set is the data the model learns from — it updates input and correct output, and the model adjusts its parameters based on the data.
• •
  Validation set helps improve the model during training.
• •
  Test set is used only after training is complete — it is not used to train, only to evaluate final model performance.

Feature Engineering

The process of selecting, transforming, and creating features from raw data to improve the performance of an ML model.

• •
  Feature creation: create new features from existing data.
• •
  Feature transformation: convert data into a better format.
• •
  Feature selection: choose only the most useful features.

Unsupervised Learning

• •
  The goal is to discover inherent patterns, structures, or relationships within input data.
• •
  The machine must uncover and create groups itself — but humans still put labels on the output groups.

Clustering groups similar data points together — e.g. customer segmentation.

Association rule learning is used to discover relationships and patterns between items in a large dataset — e.g. if a customer buys X, what else are they likely to buy? (market-basket analysis).

Anomaly Detection and Semi-Supervised Learning

Anomaly detection is a machine-learning technique used to identify data points, events, or observations that are significantly different from normal behaviour — e.g. fraud detection.

Semi-supervised learning uses a small amount of labeled data and a large amount of unlabeled data to train the system. Essentially, an algorithm trained on labeled data then labels the unlabeled data.

Model Fit

If a model has poor performance, you need to look at its fit.

1. 1.
   Overfitting — performs well on training data but does not perform well on evaluation data.
2. 2.
   Underfitting — performs poorly even on training data; could be a problem of a model that is too simple or has poor data features.
3. 3.
   Balanced — neither overfitting nor underfitting; the ideal state.

Bias and Variance

Bias:

• •
  The difference (error) between predicted and actual values.
• •
  Occurs due to a wrong choice in the ML model.

High bias: the model doesn't clearly understand the training data — considered underfitting.

Reducing bias: use a more complex model; increase the number of features.

Variance:

• •
  How much the performance of a model changes if trained on a different dataset with a same/similar distribution.

High variance: the model is very sensitive to changes in the training set — this is overfitting.

Reducing variance: feature selection; split data into training and test datasets.

Confusion Matrix and Evaluation Metrics

A confusion matrix compares predicted values against actual values for a classification problem (e.g., spam vs. not spam).

Key metrics:

• •
  Precision = TP / (TP + FP).
• •
  Recall = TP / (TP + FN).
• •
  F1 = 2 × (Precision × Recall) / (Precision + Recall).
• •
  Accuracy = (TP + TN) / (TP + TN + FP + FN).

AUC–ROC: value ranges from 0 to 1 — the larger the area under the curve, the better the classifier.

Machine Learning — Inferencing

Inferencing is when a model makes predictions on new data.

Phases of machine learning:

1. 1.
   Define business goal.
2. 2.
   ML problem framing.
3. 3.
   Data processing.
4. 4.
   Model development.
5. 5.
   Retrain.
6. 6.
   Deployment.
7. 7.
   Monitoring.
8. 8.
   Iterations.

Hyperparameter Tuning

• •
  Settings that define the model structure and the learning algorithm/process.
• •
  Set before training begins.

Tuning means finding the best hyperparameter values to optimize model performance. It improves accuracy, reduces overfitting, and enhances generalization.

• •
  Learning rate — controls how much the model updates itself after each learning step. Too high → may miss the optimal solution; too low → training becomes very slow.
• •
  Number of epochs — how many complete passes through the training dataset. Too few → underfitting; too many → overfitting.
• •
  Batch size — number of training examples processed at once.
• •
  Regularization — adjusts the balance between a simple and a complex model; helps reduce overfitting.

What to Do if Overfitting

Overfitting is when the model gives good predictions for training data but not for new data.

It occurs due to:

• •
  Training data size is too small and does not represent all possible input values.
• •
  The model trains too long on a single sample set of data.
• •
  Model complexity is high and it learns from the "noise" within the training data.

How to prevent overfitting:

• •
  Increase the training data size.
• •
  Early stopping of the model's training.
• •
  Data augmentation (to increase diversity in the dataset).
• •
  Adjust hyperparameters (but you can't "add" them).
• •
  Ensembling (combine multiple models to get accurate results).

AWS Managed AI Services — Overview

The exam expects instant recall of which pre-trained AWS AI service solves which problem. The sections that follow cover each service in turn.

• •
  AWS AI services are pre-trained ML services for your use case.
• •
  Provide responsiveness and availability.
• •
  Token-based pricing.
• •
  Specialized CPUs and GPUs for specific use cases for cost-saving.
• •
  Deployed across multiple Availability Zones and regions.

Amazon Comprehend

• •
  For Natural Language Processing (NLP).
• •
  Fully managed, end-to-end service.
• •
  Uses ML to find insights and relationships in text.

  • •
    Detects the language of the text.
  • •
    Extracts key phrases, places, people, brands, or events.
  • •
    Understands how positive or negative the text is (sentiment).

Comprehend — Custom Classification:

• •
  Organize documents into categories that you define.
• •
  Supports different document types.
• •
  Real-time analysis → single document, synchronous.
• •
  Multi-document (batch), asynchronous.

Training Data → S3 → Comprehend Custom Classification

Named Entity Recognition (NER) extracts pre-defined, grouped-phrase entities like people, places, organizations, dates, and job titles.

Custom Entity Recognition:

• •
  Analyze text for specific terms and noun-based phrases.
• •
  Extract terms like policy numbers or specialized phrases.
• •
  Train the model with custom data.

Amazon Translate and Amazon Transcribe

Amazon Translate:

• •
  Natural and accurate language translation.
• •
  Converts one language into another (e.g. English → French, French → Spanish, Spanish → Hindi).
• •
  Text-to-text translation.

Amazon Transcribe:

• •
  Automatically converts speech to text.
• •
  Uses a deep-learning model called automatic speech recognition (ASR) to convert speech into text.
• •
  Automatically removes personally identifiable information (PII) using redaction; supports toxicity detection.

Improving Transcribe accuracy — allow Transcribe to capture domain-specific or non-standard terms using two mechanisms:

• •
  Custom Vocabulary (for words) — add specific words, phrases, or technical terms; good for brand names, product names, industry-specific terms, acronyms, and abbreviations. Increases recognition of new words by providing hints (pronunciation, spelling, examples).
• •
  Custom Language Model (CLM) (for context) — train on your own domain-specific text data; good for transcription of large volumes of domain-specific speech; learns the context associated with a given word and improves accuracy for industry-specific content.

Amazon Polly and Amazon Rekognition

Amazon Polly:

• •
  Turns text into life-like speech using deep learning, allowing you to create applications that talk.
• •
  Lexicons — define how to read or pronounce specific words within a text.
• •
  SSML (Speech Synthesis Markup Language) — markup for your text to indicate how to pronounce it.

Amazon Rekognition:

• •
  Find objects, people, text, and scenes in images and video using ML.
• •
  Facial analysis and facial search to do user verification and people counting.
• •
  Use cases: labeling, content moderation, text detection, celebrity recognition, pathing, and face search and verification.

Rekognition — Custom Labels:

• •
  Label your training images and upload them to Amazon Rekognition.
• •
  Only need a few hundred images or fewer.
• •
  Create a custom model on your image data.

Labelled Images → S3 → Rekognition (Custom Labels) → Custom Classification

Rekognition — Content Moderation:

• •
  Automatically detect inappropriate, unwanted, or offensive content.
• •
  Filters out harmful images in social media, broadcast media, advertising, and more.
• •
  Brings down the human review volume to 5–8% of total content.

Amazon Lex, Personalize, and Textract

Amazon Lex:

• •
  Build a chatbot quickly for your application using voice and text.
• •
  Supports multiple languages.
• •
  The bot automatically understands the user intent to invoke the correct Lambda function.

Amazon Personalize (retail stores, media, etc.):

• •
  Fully managed ML service to build apps with real-time personalized recommendations.
• •
  Personalized product recommendations and direct marketing.
• •
  Integrates into websites, applications, SMS, and email.
• •
  Recipes — algorithms built and prepared for specific use cases. You must provide training data or target recipes.

Amazon Textract:

• •
  Automatically extract text, handwriting, and data from any scanned document.
• •
  Extract data from forms and tables.
• •
  Read and process any type of document.

Amazon Kendra, Mechanical Turk, and Augmented AI

Amazon Kendra:

• •
  Fully managed document search service powered by ML.
• •
  Extract answers from within a document.
• •
  Learns from user interaction/feedback to promote preferred results.
• •
  Ability to manually fine-tune search results.

Amazon Mechanical Turk:

• •
  Crowdsourcing marketplace to perform simple human tasks.
• •
  A distributed virtual workforce.
• •
  Use cases: image classification, data collection, and business processing.

Amazon Augmented AI (A2I):

• •
  Human oversight of ML predictions in production.
• •
  Reviewers can be your own employees or external contractors from AWS or Mechanical Turk.
• •
  The ML model can be built on AWS or elsewhere.

Medical AI Services

Amazon Transcribe Medical:

• •
  Automatically convert medical-related speech to text.
• •
  Ability to transcribe medical terminology.
• •
  Supports both real-time and batch transcriptions.

AWS Comprehend Medical detects and returns useful information in unstructured clinical text — physician's notes, test results, and discharge summaries.

• •
  Uses NLP to detect Protected Health Information (PHI).
• •
  Analyzes real-time data with Kinesis Data Firehose and stores data in S3.

AWS HealthScribe:

• •
  A HIPAA-eligible service that automatically generates clinical notes by analyzing patient–clinician conversations.
• •
  Helps you build clinical apps that provide dictation transcripts and identify speaker roles.

AI Hardware — EC2, Trainium, and Inferentia

Amazon EC2 for AI — GPU-based EC2 instances (P3, P4, P5, G3, G4, G6, and more).

AWS Trainium:

• •
  An ML chip built to perform deep learning on models with 100B+ parameters.
• •
  Around 50% cost reduction when training a model.

AWS Inferentia:

• •
  An ML chip built to deliver inference at high performance and low cost.
• •
  Up to 4× throughput and 70% cost reduction.

Amazon SageMaker — Overview

• •
  Fully managed service for developers to build ML models.
• •
  An end-to-end ML service:

  • •
    Collect and prepare data.
  • •
    Build and train the machine-learning model.
  • •
    Deploy the model and monitor the performance of predictions.

Built-in algorithms:

• •
  Supervised algorithms — linear regression and classification; KNN algorithms.
• •
  Unsupervised algorithms — K-means; anomaly detection.
• •
  Textual algorithms — NLP, summarization.
• •
  Image processing — classification, detection.

Automatic Model Tuning (AMT):

• •
  Define an objective metric.
• •
  Saves time and money.
• •
  AMT automatically chooses hyperparameter ranges, search strategy, tuning-job settings, and early-stop conditions.

SageMaker Model Deployment and Inference

• •
  Deploy with or without automatic scaling — no server to manage.
• •
  Reduces overhead.
• •
  Real-time → one prediction at a time.
• •
  Serverless.
• •
  Asynchronous → for large payload sizes up to 1 GB, with hours of processing time and near-real-time latency requirements.
• •
  Batch → pay only per use in S3; prediction for your entire dataset.

Inference types compared:

SageMaker Studio and Data Tools

SageMaker Studio:

• •
  End-to-end ML development from a unified interface.
• •
  Team collaboration.
• •
  Train and debug ML models.
• •
  Deploy ML models.
• •
  Automated workflows.

Data Wrangler:

• •
  Prepare tabular and image data for machine learning.
• •
  SQL support — a data-quality tool.
• •
  Data preparation, transformation, and feature engineering.

ML Feature Store — a feature store provides input to the ML model; used during training and for inference.

SageMaker Clarify and Explainability

SageMaker Clarify:

• •
  Evaluate foundation models.
• •
  Built-in metrics and algorithms.
• •
  Use a built-in dataset or bring your own dataset.
• •
  Evaluate human-judgment factors such as friendliness or humor.

Model Explainability:

• •
  A set of tools to help explain how ML models make predictions.
• •
  Debug predictions provided by the model after it's deployed.
• •
  Helps increase trust and understanding of the model.

Detect Bias (Clarify):

• •
  Ability to detect and explain biases in your dataset and model.
• •
  Measure bias using statistical metrics.

SageMaker Ground Truth and ML Governance

SageMaker Ground Truth:

• •
  RLHF — Reinforcement Learning from Human Feedback.
• •
  Human feedback → model review → customization and evolution; align the model to human preferences.

ML Governance:

• •
  Model Card — essential model information.
• •
  Role Manager — define roles for resources.
• •
  Model Dashboard — a centralized portal where you can view, search, and explore all your models; helps find models that violate the thresholds you set for data quality, model quality, bias, and explainability.
• •
  Model Monitor — monitors the quality of your model in production.
• •
  Model Registry — a centralized repository that lets you track, manage, and version ML models.

SageMaker Pipelines, JumpStart, and Canvas

SageMaker Pipelines:

• •
  A workflow that automates building, training, and deploying ML models.
• •
  CI/CD for ML.
• •
  Helps build, train, test, and deploy models automatically and reduce errors and costs.

SageMaker JumpStart:

• •
  An ML hub to find pre-trained foundation models, computer-vision models, or NLP models.
• •
  A large collection of models from Meta, Hugging Face, and others.
• •
  Fully customizable for your data and use case.

SageMaker Canvas:

• •
  Build ML models using a visual interface (no coding).
• •
  Allows you to use ready-to-use models from data sources on JumpStart.

Part 2: Fundamentals of Generative AI (24%)

This domain covers what generative AI is, the concepts behind it (foundation models, tokens, embeddings, context windows), and the AWS technologies for building generative-AI applications — Amazon Bedrock, Amazon Nova, Amazon Q, and PartyRock.

Generative AI and Foundation Models

• •
  A subset of deep learning.
• •
  Used to generate new data that is similar to the data it was trained on.

Examples of generated content: text, audio, video, code, and more.

The typical generative pipeline moves from raw data to a reusable base model, which is then adapted for many downstream tasks:

Unlabeled Data → Pre-train → Foundation Model → Adapt → Applications

Downstream applications include: text generation & summarization, image generation, question answering, and many others.

Foundation Model (FM):

• •
  To generate data, generative AI relies on a foundation model.
• •
  Foundation models are trained on a wide variety of input data.
• •
  Foundation models are the base for many AI applications.

Large Language Model (LLM):

• •
  Type of AI designed to generate coherent human-like text.
• •
  Trained on a large corpus of text data.
• •
  Can perform language-related tasks — translation, summarization, question answering, and content creation.

Generative AI — Core Concepts

1) Tokenization:

• •
  Breaking text into smaller units called tokens.
• •
  Word-based tokenization: text is split into individual words.
• •
  Subword tokenization: words are split further (e.g., "playing" becomes "play" + "ing").

Text → Token → Token ID → Model Processing

2) Context Window:

• •
  The amount of information an AI model can consider at one time.
• •
  A larger context window allows more information and greater coherence.

Trade-off: a larger context window needs more memory but delivers more reasoning power. It is a key factor to consider when choosing a model.

3) Embeddings:

• •
  Create a vector out of text, images, or audio, e.g. [0.21, -0.32, 0.78, …].
• •
  Embeddings enable AI to power search applications, letting models find meaning and similarity.

Amazon Bedrock — Overview

• •
  Build new AI applications on AWS.
• •
  Fully managed service — no servers for you to manage (serverless for you).
• •
  Keep control of your data and use it to train/customize the model.
• •
  Pay as you use — pay only for what you consume.
• •
  Unified API — one API connects you to the full Bedrock model catalogue (one API for many models).
• •
  Security, Privacy & Governance are built-in Responsible AI features.

Amazon Bedrock works on a private copy of the foundation model, available only to you, which you can further fine-tune with your own data. Your own data is then used to specialize the model:

Your Data → Foundation Model → Trained / Customized Model

Amazon Nova

• •
  A family of AI foundation models built by AWS.
• •
  Designed to be fast, cost-effective, and enterprise-ready.
• •
  Accessible through Amazon Bedrock.

Request/response flow: Applications → Amazon Bedrock → Amazon Nova, and back again.

Specialized Nova models:

• •
  Nova Canvas — state-of-the-art image generation model.
• •
  Nova Reel — state-of-the-art video generation model.
• •
  Nova Sonic — conversational speech understanding and generation in multiple languages.

Nova 2:

• •
  A new model with enhanced capabilities to build complex AI apps.
• •
  Interactive chatbots, document and video analysis, and AI agent creation.

Multimodal model: an AI system capable of processing multiple types of input data such as text, image, and audio, enabling use across different domains.

Amazon Bedrock — Pricing

1) On-Demand Pricing:

• •
  Pay as you go.
• •
  Charged on tokens — both input tokens and output tokens.

2) Batch Processing:

• •
  Millions of prompts are processed offline for lower cost.
• •
  Suited to large workloads.
• •
  Up to 50% discounts.

3) Provisioned Throughput:

• •
  Purchase reserved model capacity for a minimum time period.
• •
  Works with base, fine-tuned, and custom models.

Amazon Q Business

• •
  A fully managed generative-AI assistant for your employees.
• •
  Based on your company's knowledge and data.
• •
  Built on Amazon Bedrock (using multiple foundation models).

Core capabilities:

• •
  Enterprise knowledge search.
• •
  Generative AI responses.
• •
  Role-based access control.
• •
  Document summarization.
• •
  Task assistance.

Amazon Q Business + IAM Identity Center:

• •
  Users can be authenticated through IAM Identity Center.
• •
  Users receive responses grounded only in the documents they have access to.

Admin Control:

• •
  Control and customize responses to your organization's needs.
• •
  Admin control acts as guardrails.
• •
  Block specific words or topics.
• •
  Respond only with internal information.

Amazon Q Developer (Kiro)

• •
  Answers questions about AWS documentation and AWS service selection.
• •
  Answers questions about the resources in your AWS account.
• •
  Helps you with bill analysis, error resolution, and troubleshooting.
• •
  An AI coding companion (Kiro) to help you build new applications.
• •
  Supports Java, Python, JavaScript, C#, and more.
• •
  Provides real-time code suggestions and security scans.

PartyRock

• •
  Not an AWS service (it is a standalone playground).
• •
  Runs an AI app-building application (powered by Amazon Bedrock).
• •
  Allows you to experiment with creating generative AI apps using various foundation models.
• •
  No coding required.
• •
  The UI is similar to the Amazon Q app.

Part 3: Applications of Foundation Models (28%)

The heaviest domain. It covers the full journey of applying a foundation model: choosing one, prompting it well, tuning its inference parameters, grounding it with RAG, extending it with agents, customizing it through fine-tuning and distillation, and evaluating the result.

Choosing a Base Model

When selecting a foundation (base) model, weigh several factors against your use case:

• •
  Model type, performance, capabilities, constraints, and compliance.
• •
  Level of customization, model size, inference options, licensing agreements, context window, and latency.
• •
  Multi-modal support (models that handle more than one data type).

Amazon Titan:

• •
  A high-performing foundation model built by AWS.
• •
  Image, text, and multimodal model choices, available via fully managed APIs.
• •
  Can be customized with your own data.

Rule of thumb: smaller models are generally more cost-effective.

Customizing a base model on your own data produces a specialized custom model:

Our Data → Foundation Model → Custom Model

Benefits of a custom model: better results, higher accuracy, and lower latency.

Prompt Engineering

The process of developing, designing, and optimizing prompts to improve the quality of a model's response.

The four elements of a good prompt: Instruction, Context, Input data, and Output indicator.

1) Zero-Shot Prompting:

• •
  The model is given a task without any examples.
• •
  Relies entirely on its pre-trained knowledge.

2) Few-Shot Prompting:

• •
  Provide one or more examples before asking the model to perform the task.

3) Chain-of-Thought Prompting:

• •
  Break a problem into reasoning steps so the model works through it logically.

4) Retrieval-Augmented Generation (RAG):

• •
  Combines the model with external knowledge sources for grounded answers.

Good prompts pair clear instructions with relevant context and good input data; iterate and refine prompts to continuously improve results.

Prompt Response Optimization (Inference Parameters)

These inference parameters shape how a model produces its output:

• •
  System prompts — define how the model should behave.
• •
  Temperature — controls the creativity of the model's output.

  • •
    Low temperature → outputs are more conservative, predictable, and focused on the most likely response.
  • •
    High temperature → outputs are more diverse, creative, and unpredictable (and might be incorrect).
• •
  Top P — controls the pool of candidate words by cumulative probability.

  • •
    Low P → considers only the most likely words → more correct responses.
  • •
    High P → considers a broad range of possible words → more creative, diverse output.
• •
  Length — controls the maximum length of the answer.
• •
  Stop sequence — tokens that signal the model to stop generating output.
• •
  Top K — limits the number of probable words considered.

  • •
    Low K → more correct responses (fewer, more probable words).
  • •
    High K → more probable words considered → more diverse and creative output.

Tune these settings based on your use case and goals: length and stop sequences manage output size, while system prompts define behaviour.

RAG and the Knowledge Base

RAG = Retrieval-Augmented Generation.

• •
  Allows a foundation model to reference a data source outside of its training data.
• •
  Bedrock takes care of creating vector embeddings in a database of your choice, based on your data (documents).
• •
  Use where real-time data needs to be fed into the FM.

How a RAG query flows:

1. 1.
   Store the document(s).
2. 2.
   User asks a question.
3. 3.
   Retrieve relevant information [Docs + Related Docs].
4. 4.
   Send context to the LLM [Query + Related Docs].
5. 5.
   The FM generates the answer.

RAG vector databases on AWS:

RAG data sources include: Amazon S3, OneDrive, Microsoft SharePoint, Salesforce, web pages — and more over time.

Use cases: customer-service chatbots, legal research and analytics, and healthcare question answering.

AWS offers scalable, secure, and cost-effective vector-store options — choose the right one based on your data and use case.

Amazon Bedrock — Agents

• •
  A fully managed capability that helps you build AI applications that can act.

An agent can:

• •
  Understand user requests.
• •
  Break tasks into steps.
• •
  Invoke APIs and AWS Lambda functions.
• •
  Access knowledge bases.
• •
  Execute workflows automatically.
• •
  Perform tasks in the correct order and ensure information is passed correctly between tasks.

Agents are configured to perform specific, pre-defined action groups.

Bedrock Agents + CloudWatch:

• •
  Monitor, troubleshoot, and analyze agent behaviour.
• •
  Can include text, images, and embeddings.
• •
  Published metrics flow from Bedrock to CloudWatch.
• •
  You can build CloudWatch alarms on these metrics.

Fine-Tuning a Model

• •
  Adapt a copy of a foundation model with your own data.
• •
  Fine-tuning changes the weights of the base foundation model.
• •
  Training data must adhere to a specific format and be stored in Amazon S3.

Supervised Fine-Tuning (labeled data):

• •
  Improves the performance of the model on specific tasks.
• •
  Further trains the model on a particular field or area of knowledge.
• •
  Uses input–output pairs — a prompt (input) and its provided response (output).

Reinforcement Fine-Tuning (feedback based):

• •
  Improves your foundation model using feedback-based learning.
• •
  You provide input data (prompts).
• •
  A reward function evaluates the model's outputs and judges which response is best.

Example: for a single prompt the model generates multiple candidate outputs (Output 1…Output n); each is scored, and the reward signal steers the model toward higher-scoring responses.

Distillation

• •
  Makes a model smaller and faster.
• •
  Up to 75% reduction in size versus the original model.
• •
  A larger "teacher" model transfers its knowledge to a smaller "student" model.
• •
  Focuses on efficiency, speed, and cost reduction.

The knowledge-transfer flow:

Teacher (large model) → Knowledge Transfer → Student (small model)

Evaluating a Model

1) Automatic Evaluation:

• •
  The model is evaluated using metrics and benchmarks.

An expected output is compared against the model output for a given question: correct output raises the score; wrong output lowers it. Common metrics include Accuracy, BLEU, ROUGE, F1 score, and more.

2) Human Evaluation:

• •
  Human reviewers assess the model's output.

Reviewers rate: accuracy, helpfulness, safety, tone, and relevance.

Business metrics to evaluate a model on:

• •
  User satisfaction.
• •
  Average revenue per user.
• •
  Cross-domain performance.
• •
  Efficiency.
• •
  Conversion rate.

Part 4: Guidelines for Responsible AI (14%)

This domain covers the dimensions of responsible AI, the AWS tooling that supports them (including Bedrock Guardrails), and the risks unique to generative AI — from toxicity and hallucinations to prompt attacks and jailbreaking.

Responsible AI — Core Dimensions

Responsible AI sits at the intersection of people, process, and technology. There are seven (plus) core dimensions to understand deeply for the exam.

AI Solution Lifecycle: Data Collection → Model Training → Testing & Validation → Deployment → Monitoring & Feedback, with continuous improvement looping back.

Connect each dimension with the AWS service that supports it, and focus on use cases and best practices — it's not just about accuracy, it's about impact.

Responsible-AI AWS Services

• •
  Amazon Bedrock — human and automatic model evaluation; worldwide filters for blogs, vulnerable topics, harmful content, and more.
• •
  SageMaker Clarify — detect bias in data and explain model predictions.
• •
  SageMaker Data Wrangler — data prep, cleaning, and visualization.
• •
  SageMaker Model Monitor — monitor the model in production.
• •
  Amazon Augmented AI (A2I) — human + AI collaboration.
• •
  Amazon Forecast — time-series forecasting.

Amazon Bedrock — Guardrails

• •
  Control the interaction between users and foundation models (FMs).
• •
  Filter undesirable and harmful content.
• •
  Enhance privacy.
• •
  Support PII (personally identifiable information) handling.
• •
  Reduce hallucinations.
• •
  Create multiple guardrails and versions, and analyze user input that may violate guardrails.

How it works: a user prompt passes through Bedrock Guardrails. Allowed requests reach the FM and a response returns; blocked or filtered content is stopped and returned to the user as a safe message.

Guardrails protect users and your business by filtering harmful content, help maintain privacy and comply with policies (e.g., PII support), reduce hallucinations, and let you monitor and analyze inputs to continuously improve.

Interpretability — Decision Trees and PDP

High interpretability — Decision Tree:

• •
  A supervised-learning algorithm used for classification and regression tasks.
• •
  Splits data into branches based on feature values.
• •
  Can overfit if you have too many branches.
• •
  Easy to interpret and gives a clear visual representation.

Partial Dependence Plots (PDP):

• •
  Show how a single feature can influence the predicted outcome while holding other features constant.
• •
  Help with interpretability and explainability.

Human-Centered Design (HCD) for Explainable AI

• •
  An approach to design AI systems with priorities for human needs.
• •
  Design for amplified decision-making — minimize risk and errors in critical or high-pressure environments; design for clarity, simplicity, and usability.
• •
  Design for unbiased decision-making — the decision process should be free from bias; train decision-makers to recognize and remove biases.
• •
  Design for human and AI teaming — personalization, user-centered design, and AI systems that learn from human instructors and experts.

Generative-AI Risks — Toxicity and Hallucinations

1) Toxicity:

• •
  Providing content that is offensive, disturbing, or inappropriate.
• •
  Examples: hate speech, insulting words, rude content, and harassment.

Mitigation: create training data by identifying and removing offensive phrases in advance; use a guardrail model to detect and filter out unwanted content.

2) Hallucinations:

• •
  Assertions or claims that sound true but are incorrect.

Cause: next-word probability sampling employed by LLMs — hard to control content that may not exist, even though it seems plausible.

Mitigation: educate users that model-generated content must be checked; ensure verification of content with independent sources.

Plagiarism, Cheating, and Prompt Attacks

Plagiarism and cheating:

• •
  Concerns that generative AI can be used to write college essays, job-application samples, and other forms of cheating or copying.
• •
  Use technologies to detect if text or images have been generated with AI.

Prompt attack methods:

• •
  Poisoning — intentional introduction of malicious or biased data into the training dataset of a model; leads to the model producing biased, offensive, or harmful outputs.
• •
  Misleading Prompt Injection — influences the output by embedding specific instructions within the prompts themselves; hijacks the model behaviour and makes it produce output that aligns with the attacker's intentions.

Data Risks — Exposure, Leakage, and Jailbreaking

Exposure:

• •
  The risk of exposing sensitive or confidential information to the model during training or inference.
• •
  The model can then reveal this sensitive data.

Prompt Leakage:

• •
  The unintentional disclosure or leakage of the prompt or input used within a model.
• •
  It can expose protected data or other data used by the model, such as how the model works.

Jailbreaking:

• •
  A model is typically trained with certain ethical and safety constraints in place to prevent unsafe or harmful outputs.
• •
  Jailbreaking prompts try to bypass these constraints, which leads to unsafe or harmful outputs.

Part 5: Security, Compliance, and Governance for AI Solutions (14%)

This domain covers how responsibility is shared between you and AWS, how AI workloads are governed and audited, and the core AWS services — identity, storage, compute, and security tooling — that every AI solution rests on.

The Shared Responsibility Model

Security in the AWS cloud is a partnership between you and AWS. Responsibilities are split cleanly between the two parties.

• •
  AWS responsibility → security "of" the cloud.
• •
  Customer responsibility → security "in" the cloud.

Customer — security "IN" the cloud (responsible for securing your data, applications, access, and configurations):

• •
  Customer data.
• •
  Platforms, applications, and IAM (identity & access management).
• •
  Operating system, network, and firewall configuration.
• •
  Encryption of data at rest and in transit.

AWS — security "OF" the cloud (protects the global infrastructure that runs AWS services):

• •
  Software running the managed services.
• •
  Compute, storage, database, and networking infrastructure.
• •
  Hardware and the global infrastructure.
• •
  Regions, Edge locations, and Availability Zones (AZ).

Shared control areas:

• •
  Patch management — keeping software and systems updated.
• •
  Configuration management — managing settings and configurations securely.
• •
  Awareness and training — user awareness and security best practices.

Regulated Workloads and Model Cards

Some industries require an extra level of compliance: financial services, healthcare, and aerospace.

AI standard compliance challenges:

• •
  Complexity and opacity.
• •
  Dynamism and adaptability.
• •
  Emerged capabilities.
• •
  Unique risks — algorithmic bias and human bias.
• •
  Algorithm accountability.

Model Cards:

• •
  A standardized format for documenting the key details about an ML model.
• •
  Details about the dataset used, its source, licensing, and any known biases or quality issues in the training data.
• •
  Helps support audit activities.

Importance of Governance and Compliance

• •
  Managing, optimizing, and scaling organizational AI initiatives.
• •
  Governance is instrumental to building trust.
• •
  Ensures responsible and trustworthy AI solutions.
• •
  Mitigates risk — bias, privacy violation, and unintended consequences.
• •
  Establishes clear policies, guidelines, and oversight mechanisms so AI systems align with legal and regulatory requirements.
• •
  Protects from potential legal and reputational risks.

Governance framework:

• •
  Establish an AI governance board or committee.
• •
  Define roles and responsibilities.
• •
  Implement policies and procedures.

AWS governance & compliance services:

Governance and Data Management Strategies

Governance strategies:

• •
  Policies → principles, guidelines, and responsible AI.
• •
  Review (ad-hoc) → combination of technical, legal, and responsible-AI review.
• •
  Review strategies.
• •
  Transparency standards — document your training requirements.

Data governance strategies:

• •
  Responsible AI.
• •
  Governance structure and roles.
• •
  Data sharing and collaboration.

Data management concepts:

• •
  Data lifecycle, data logging, data residency, data monitoring, data analysis, and data retention.

Data Lineage

• •
  Source citation — attributes and acknowledges the sources of data (datasets, databases, other sources) along with relevant licenses, terms of use, or permissions.
• •
  Documenting data origins — details of the collection process, methods used to clean and import the data, and pre-processing and transformation applied.
• •
  Cataloging — organization and documentation of datasets; helps with transparency, traceability, and accountability.
• •
  Data lineage — track the flow of data from source to final output; helps identify issues, dependencies, and transformations; supports audit activities.

Security and Monitoring for AI Systems

Security and privacy for AI systems:

• •
  Threat detection.
• •
  Vulnerability management.
• •
  Infrastructure protection.
• •
  Prompt injection defense.
• •
  Data encryption.

Monitoring AI systems:

• •
  Performance metrics — model accuracy, precision, recall, F1-score, and latency.
• •
  Infrastructure monitoring.
• •
  Bias and fairness.
• •
  Compliance and responsible AI.

AWS Security and Foundational Services

1) IAM — Users and Groups (Global service):

• •
  Identity and Access Management.
• •
  Users are people within your organization and can be grouped.
• •
  Groups only contain users.
• •
  A user can belong to multiple groups.
• •
  IAM permissions → users or groups can be assigned a JSON document called a policy.

Roles for services: assign permissions to AWS services with IAM roles.

2) Amazon S3:

• •
  Allows people to store objects in buckets.
• •
  Buckets are defined at the region level.
• •
  Use cases: backup and storage, disaster recovery, archive, media hosting, static website, and software delivery.
• •
  Objects → each file has a key. Max object size is 5 TB. Use multi-part upload if the size is more than 5 GB.

Other AWS services:

S3 Storage Classes

• •
  Durability — high durability of objects; the same for all storage classes. S3 is designed for 11 nines (99.999999999%) durability.
• •
  Availability — measures how readily available a service is; S3 Standard is 99.99% available.

Amazon EC2 and AWS Lambda

Amazon EC2 — Elastic Compute Cloud = IaaS:

• •
  Provides virtual machines.
• •
  Stores data on EBS volumes.
• •
  Distributes load across multiple instances.
• •
  Scales services using an Auto-Scaling Group.
• •
  Uses data → bootstrapping the instance using an EC2 user-data script.
• •
  Runs a launching command line or web-server hosts.

AWS Lambda:

• •
  Virtual function → no servers to manage.
• •
  Limited by time → short execution.
• •
  Run on demand → executes only when triggered.
• •
  Scaling is automated → AWS automatically scales the function based on demand.

Lambda features: cost efficient (pay only for the time your code runs), highly available (built-in high availability and fault tolerance), and easy integration (works with many AWS services).

Lambda use cases: event processing (S3, DynamoDB, etc.), scheduled tasks (e.g. daily), and backend for APIs.

AWS Macie, Config, and Inspector

AWS Macie:

• •
  Fully managed data-discovery and data-privacy service that uses ML and pattern matching to discover and protect your sensitive data in AWS.
• •
  Helps identify and alert you to sensitive data, such as PII.

S3 (Analyze) → Macie (Notify) → Lambda (Integration) → Amazon EventBridge

AWS Config:

• •
  Helps with auditing and recording compliance.
• •
  Helps record configuration and changes over time.
• •
  You can receive alerts for any changes (via SNS).

AWS Inspector:

• •
  Automated security assessment service.
• •
  Regularly and continuously scans your AWS resources and workloads.
• •
  Reports and integrates with AWS Security Hub.
• •
  For EC2 instances, ECR, and Lambda.

Continuous scanning of infrastructure only when needed: patch vulnerabilities and network reachability.

AWS CloudTrail, Audit Manager, and Trusted Advisor

AWS CloudTrail:

• •
  Provides governance, compliance, and audit for your AWS account.
• •
  Enabled by default.
• •
  Logs all API calls made within your AWS account.

AWS Audit Manager:

• •
  Assess risk and compliance of your AWS workloads.
• •
  Continuously audit AWS resources using audit checks and reports.

Trusted Advisor:

• •
  Analyzes your AWS account and provides recommendations.

6 Pillars of Excellence: Operational Excellence, Cost Optimization, Performance, Security, Fault Tolerance, and Service Quotas.

AWS Artifact provides customers with on-demand access to AWS compliance documentation and AWS agreements.

Final Review and Exam Strategy

You've reached the end of the study material. Before your final revision pass, confirm you can speak confidently to every topic on this list:

• •
  AI / ML concepts and terminologies, and the ML solution lifecycle.
• •
  Amazon Bedrock and the types of inference on AWS.
• •
  Amazon Comprehend, Rekognition, Textract, and Translate.
• •
  Amazon Polly, Lex, Kendra, and Forecast.
• •
  Amazon SageMaker and its key components.
• •
  Responsible AI principles, and AI governance and compliance.
• •
  Security, privacy, and best practices, plus cost optimization for AI/ML.
• •
  Real-world use cases and examples.

These notes were designed to help you understand AWS AI services, concepts, and best practices in a simple and effective way. Learning never stops — the more you practice, the more confident you become. Keep learning, keep building, and keep exploring the power of AI on AWS. All the best for your certification journey and beyond!

Exam-Day Checklist

Domain 1 — Fundamentals of AI and ML (20%)

• •
  AI ⊃ ML ⊃ Deep Learning ⊃ Generative AI; all GenAI is DL, not all DL is GenAI.
• •
  Labeled data = supervised (regression = continuous, classification = discrete); unlabeled = unsupervised (clustering, association rules); a little of both = semi-supervised.
• •
  Overfitting = great on training, poor on new data → more data, early stopping, augmentation, regularization, ensembling. Underfitting = poor everywhere → more complexity/features.
• •
  High bias = underfitting; high variance = overfitting; balanced = low bias / low variance.
• •
  Precision = TP/(TP+FP); Recall = TP/(TP+FN); F1 balances both; AUC closer to 1 is better.
• •
  Service recall: Comprehend = NLP/sentiment; Translate = text-to-text; Transcribe = speech-to-text (ASR, PII redaction); Polly = text-to-speech; Rekognition = images/video; Lex = chatbots; Textract = document extraction; Kendra = document search; Personalize = recommendations; A2I = human review.
• •
  SageMaker: real-time / serverless / asynchronous (up to 1 GB) / batch inference; Clarify = bias + explainability; Ground Truth = RLHF; JumpStart = model hub; Canvas = no-code; Pipelines = CI/CD for ML.
• •
  Trainium = training chip (~50% cost cut); Inferentia = inference chip (up to 4× throughput, 70% cost cut).

Domain 2 — Fundamentals of Generative AI (24%)

• •
  Foundation models are pre-trained on broad unlabeled data, then adapted; LLMs are FMs for human language.
• •
  Tokenization → token IDs → model; context window = how much the model considers at once; embeddings = vectors that capture meaning/similarity.
• •
  Bedrock = serverless, unified API, private model copy, your data stays yours.
• •
  Nova family: Micro (text, lowest latency) → Lite (cheap multimodal) → Pro (balanced) → Premier (most capable); Canvas = image, Reel = video, Sonic = speech.
• •
  Bedrock pricing: on-demand per token; batch up to 50% off; provisioned throughput = reserved capacity.
• •
  Q Business = employee assistant on company data (IAM Identity Center); Q Developer = AWS expert + coding companion; PartyRock = no-code Bedrock playground (not an AWS service).

Domain 3 — Applications of Foundation Models (28%)

• •
  Model choice factors: performance, cost, customization, context window, latency, licensing, multimodal support; smaller models are generally cheaper.
• •
  Prompting ladder: zero-shot → few-shot → chain-of-thought → RAG. A good prompt = instruction + context + input data + output indicator.
• •
  Low temperature/Top P/Top K → focused and accurate; high values → creative and diverse.
• •
  RAG = external knowledge, no retraining, fewer hallucinations, current data; vector stores: OpenSearch, Aurora PostgreSQL, Neptune Analytics, S3 Vectors.
• •
  Agents break tasks into steps, invoke APIs/Lambda, use knowledge bases, run action groups; monitor via CloudWatch.
• •
  Fine-tuning changes model weights (training data in S3); supervised fine-tuning is usually cheaper than reinforcement fine-tuning.
• •
  Distillation: teacher → student, up to 75% smaller, faster and cheaper inference.
• •
  Evaluation: automatic (BLEU, ROUGE, F1, accuracy) = fast and scalable; human = judgment factors; also track business metrics.

Domain 4 — Guidelines for Responsible AI (14%)

• •
  Know the core dimensions and their AWS pairings: fairness, privacy/security (Macie, KMS, IAM), robustness, governance (Model Cards), safety (Rekognition moderation), controllability, explainability (Clarify/SHAP), interpretability (decision trees, PDP).
• •
  Bedrock Guardrails filter harmful content, handle PII, and reduce hallucinations.
• •
  Risk vocabulary: toxicity, hallucination, poisoning, prompt injection, exposure, prompt leakage, jailbreaking — and one mitigation for each.
• •
  Human-centered design: amplified, unbiased, human+AI teaming decisions.

Domain 5 — Security, Compliance, and Governance (14%)

• •
  Shared responsibility: AWS = "of" the cloud; customer = "in" the cloud; shared = patching, configuration, training.
• •
  Governance toolkit: Inspector = vulnerabilities; Audit Manager = audit evidence; Artifact = compliance reports; CloudTrail = API activity log (on by default); Config = configuration change history; Trusted Advisor = best-practice checks across 6 pillars.
• •
  Macie = discover PII in S3; Model Cards + data lineage support audits and transparency.
• •
  Core services: IAM users/groups/policies/roles; S3 (5 TB max object, 11 nines durability, storage classes); EC2 = IaaS; Lambda = serverless, event-driven, short-lived.

Official Resources

• •
  AWS Certified AI Practitioner official exam page — registration, scheduling, and the latest exam policies.
• •
  AIF-C01 official exam guide (PDF) — the authoritative domain and task-statement breakdown this guide mirrors.
• •
  AWS Skill Builder — the official Exam Prep Plan for AIF-C01, including the Official Practice Exam and Official Pretest.