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

This domain asks about basic AI/ML concepts, types of learning, use cases, inference, and lifecycle. Revision 1.1 explicitly adds agentic AI, the choice between traditional ML and foundation models, and current services such as Amazon Bedrock, Amazon Q, Amazon Quick, and Kiro.

1. Basic Concepts: AI, ML, Deep Learning, GenAI, and Agentic AI#

Artificial Intelligence (AI): broad field; any technique that enables a machine to mimic human intelligent behavior (includes rule-based systems, not just learning).
Machine Learning (ML): subset of AI; the system learns patterns from data rather than explicitly programmed rules.
Deep Learning (DL): subset of ML based on neural networks with multiple layers; excels in unstructured data (images, audio, text).
Generative AI (GenAI): creates new content—text, images, audio, video, or code—based on learned patterns. Modern LLMs and other foundation models typically use deep learning.
Agentic AI: systems that pursue a goal through planning, memory, tools, and one or more agents. It is not an additional level in the hierarchy: it is an application pattern that typically uses generative models.
Exam-safe relationship: ML is part of AI and DL is part of ML. GenAI describes what the system produces; agentic AI describes how it organizes actions to achieve a goal. Avoid treating them as strictly interchangeable subsets.
Structured data (tables, CSV) vs unstructured (text, images, audio) vs semi-structured (JSON, XML). DL shines in unstructured data; classical ML is often sufficient for tabular data.

2. Types of Learning#

Type	Data	Use Cases
Supervised learning	Labeled (features + label)	Spam classification, price prediction (regression)
Unsupervised learning	Unlabeled	Customer clustering, anomaly detection, dimensionality reduction
Semi-supervised	Few labeled + many unlabeled	Classification with expensive labeling
Reinforcement learning (RL)	Agent + environment + rewards	Robotics, control, and games
Self-supervised	Labels come from the data itself	LLM pre-training (predicting the next token)

Classification (discrete label: binary or multi-class) vs regression (continuous value). If the question says "predict price / demand / a number" → regression; "classify into categories / detect yes-no fraud" → classification.
Clustering has no labels: if the scenario says "group similar customers without predefined categories" → unsupervised (K-means).

3. ML Lifecycle (ML lifecycle)#

Business problem definition (Do you even need ML? If fixed rules suffice, do not use ML).
Data collection and preparation: cleaning, handling missing values, feature engineering, train/validation/test split.
Training: algorithm selection, tuning hyperparameters (set by the human before training) vs parameters/weights (learned by the model).
Evaluation: metrics on test data the model has not seen.
Deployment: real-time endpoint vs batch inference.
Monitoring: detection of data drift and model drift; retraining.

MLOps: applying DevOps practices to the ML lifecycle (automation, CI/CD of models, data and model versioning, continuous monitoring).

Sources and Ways to Serve Models#

Open source pre-trained model: accelerates startup and allows inspection or self-hosting, but the team assumes evaluation, security, capacity, and operations.
Custom model: maximum control and specialization; requires data, compute, expertise, and maintenance.
Managed API: lower operational burden and scaling delegated to the provider.
Self-hosted API: more control over infrastructure, residency, and optimization, in exchange for operating the service.

4. Overfitting, Underfitting, and Model Quality#

Overfitting: excellent on training, poor on test → memorizes instead of generalizing. Mitigation: more data, regularization, early stopping, dropout, simpler models.
Underfitting: poor on training and on test → model too simple or insufficient data/features.
Bias-variance tradeoff: high bias ≈ underfitting; high variance ≈ overfitting.

5. Technical and Business Metrics#

Classification: confusion matrix, accuracy (misleading with imbalanced classes), precision (of what I predicted positive, how many I got right — matters when false positive is costly), recall (of the actual positives, how many I detected — matters when false negative is costly, e.g., cancer detection or fraud), F1 (harmonic mean of both), AUC-ROC.
Regression: MAE, MSE, RMSE, R².
Business: cost per user or interaction, development cost, customer satisfaction and feedback, completion rate, and ROI. A technically superior model adds no value if it worsens cost, latency, or business outcome.
Exam tip: "detect as many fraud cases as possible even if there are false alarms" → maximize recall; "avoid accusing innocent customers" → precision.

6. Inference: Modalities#

Real-time inference: low latency, always-on endpoint, sustained traffic.
Batch (transform): large volumes with no urgency; lower cost, no persistent endpoint.
Asynchronous inference: large payloads, latency-tolerant, request queue.
Serverless inference: intermittent or unpredictable traffic, pay-per-use, tolerates cold starts.

7. Choosing Traditional ML, a Foundation Model, or Rules#

Approach	Best Fit
Deterministic rules	Exact outcome, stable policy, and explainable logic; no prediction needed
Traditional ML	Tabular data or time series, bounded objective, lower cost/latency, and high explainability
Foundation model	Text, image, or open-ended tasks; generation, summarization, conversation, and cross-domain adaptation
Agentic system	Multi-step objective requiring tools, memory, decision-making, and orchestration

In regulated sectors, explainability, privacy, cost, and operational constraints may make a traditional model preferable even if an FM can solve the task.

8. AWS AI/ML Services: Which to Choose#

General exam rule: if a managed AI service exists that solves the use case, it is the answer (lowest operational effort); SageMaker AI is for when you need to build or train your own models.

Pre-trained AI Services (No ML Expertise Required)#

Amazon Comprehend: NLP on text — sentiment, entities, key phrases, PII detection, document classification.
Amazon Rekognition: vision — objects, faces, content moderation in images and video, text in images.
Amazon Textract: extract text, tables, and forms from scanned documents/PDFs (more than simple OCR: understands structure).
Amazon Transcribe: speech-to-text (audio → text).
Amazon Polly: text-to-speech (text → voice).
Amazon Translate: automatic translation between languages.
Amazon Lex: conversational chatbots with voice and text (the Alexa engine).
Amazon Kendra: intelligent enterprise search with natural language over document repositories.
Amazon Personalize: personalized recommendations (Amazon.com style) without ML expertise.
Amazon Q Business: generative enterprise assistant over company data.

Platform, Foundation Models, and Agentic Work#

Amazon Bedrock: managed API access to foundation models and capabilities to create GenAI applications and agents without managing model infrastructure.
Amazon SageMaker AI: platform to create, train, adapt, deploy, and monitor models with greater control over the ML lifecycle.
SageMaker JumpStart: catalog and starting point for pre-trained models and solutions within SageMaker AI.
Amazon Q: family of assistants for enterprise work and software development.
Amazon Quick: AI workspace to research, analyze data, visualize, and automate work with chat and agents.
Kiro: agentic development environment guided by specifications, steering, and hooks.
Strands Agents: open-source SDK to build agents; Amazon Bedrock AgentCore provides managed infrastructure to run and operate them.
Amazon Nova: family of AWS foundation models available in the Bedrock ecosystem.
AWS Transform: agentic service to modernize applications and workloads.

Human in the Loop#

Amazon A2I (Augmented AI): incorporates human review of low-confidence predictions (e.g., reviewing questionable Textract extractions).

9. Common Exam Traps#

AI vs ML vs DL: if the system uses fixed, hand-written rules, it is AI but not ML.
"No ML experience" / "least operational overhead" → managed AI service (Comprehend, Rekognition...), not SageMaker AI.
Textract vs Rekognition: documents/forms → Textract; photos/video/faces → Rekognition (although Rekognition also reads text in scene images, e.g., license plates).
Kendra vs Personalize: document search → Kendra; product/content recommendations → Personalize.
Transcribe vs Polly: they are inverses — audio→text vs text→audio. Read the conversion context carefully.
Hyperparameter vs parameter: hyperparameter is set by the human before training (learning rate, epochs); parameter (weights) is learned by the model.
Accuracy with imbalanced classes: a model that always says "no fraud" can have 99% accuracy and be useless → look at precision/recall/F1.
Inference type: "millions of records every night" → batch; "immediate response to the user" → real-time; "sporadic and unpredictable traffic" → serverless.
Traditional ML vs FM: bounded, explainable tabular prediction → Traditional ML; open-ended content generation or comprehension → FM.
Bedrock vs SageMaker AI: consuming FMs via API with low operations → Bedrock; controlling training, infrastructure, or the ML lifecycle → SageMaker AI.
Strands vs AgentCore: Strands builds the agent logic; AgentCore provides managed runtime and operational services.
When NOT to use ML?: if the problem can be solved with simple deterministic logic or there is no data, the correct answer is not to use ML.

10. Mini-Review Scenarios (Exam Format)#

"An insurer wants to predict next quarter's claim amounts with tabular data and needs to explain every decision." → Regression with Traditional ML in SageMaker AI.
"Group support tickets into topics without predefined categories." → Unsupervised learning (clustering).
"Extract amounts and fields from scanned PDF invoices." → Amazon Textract.
"Analyze the sentiment of product reviews without an ML team." → Amazon Comprehend.
"Convert company podcasts into text articles." → Amazon Transcribe.
"Give a natural voice to a telephone assistant." → Amazon Polly (+ Lex for conversation).
"Search thousands of internal documents with natural language questions." → Amazon Kendra.
"Recommend movies based on each user's history." → Amazon Personalize.
"Have a human review predictions with confidence < 80%." → Amazon A2I.
"Summarize and answer open-ended questions about contracts via a managed API." → Foundation model in Amazon Bedrock.
"Research a market, consult tools, and prepare a report in multiple steps." → Agentic system; Amazon Quick for the managed workspace or Strands + AgentCore for a custom solution.
"Always apply the same discount according to a closed legal table." → Deterministic rules, not ML.
"The model performs 99% on training and 60% on test." → Overfitting; regularization/more data/early stopping.
"The team wants to train, tune, and deploy a custom model end-to-end." → Amazon SageMaker AI.

11. Quick Glossary#

Term	Exam Definition
Feature	Model input variable
Label	Target value to be predicted (supervised)
Feature engineering	Creating/transforming features to improve the model
Training set	Data used for the model to learn
Validation set	Data to adjust hyperparameters
Test set	Unseen data to evaluate generalization
Epoch	One complete pass through the training set
Learning rate	Hyperparameter: size of the update step
Inference	Using the trained model to predict
Data drift	Production data changes relative to training data
Model drift	Model quality degrades over time
Confusion matrix	TP/FP/TN/FN table of a classifier
Ensemble	Combining multiple models to improve the result
Transfer learning	Reusing a pre-trained model on a new task
MLOps	DevOps applied to the ML lifecycle

12. Pre-exam Checklist#

I explain the relationship between AI, ML, and DL, and why GenAI and agentic AI describe distinct capabilities or patterns.
I distinguish supervised / unsupervised / semi-supervised / RL with an example of each.
I distinguish classification vs regression vs clustering based on the problem statement.
I know when to choose precision and when to choose recall (cost of FP vs FN).
I explain overfitting vs underfitting and how to mitigate them.
I know what each managed AI service does within scope (Comprehend, Rekognition, Textract, Transcribe, Polly, Translate, Lex, Kendra, and Personalize).
I choose between traditional ML and foundation models based on task, explainability, cost, and operations.
I distinguish between Bedrock, SageMaker AI, JumpStart, Amazon Q, Amazon Quick, Kiro, Strands Agents, and AgentCore.
I correctly choose between real-time, batch, async, and serverless inference.
I remember that "least operational overhead" points to the managed service.

Mapping to the repo#

This domain corresponds to Module 1 (LLM Fundamentals and APIs) and the evaluation section of Module 2.

Verified Official Sources#

⚠️ Note: Content, weights, and services may change. This guide was verified on August 21, 2026; check the official documentation before taking the exam.