TL;DR
Certification earned
I organized these notes around the official AIF-C01 domains, with core concepts, AWS services, and practice questions to help others prepare. The questions are study material, not a complete exam simulation.
Exam Overview and Domain Breakdown
The AWS Certified AI Practitioner exam focuses on five domains:
| Domain | Domain Description | Exam Weighting |
|---|---|---|
| Domain 1 | Fundamentals of AI and Machine Learning (ML) | 20% |
| Domain 2 | Fundamentals of Generative AI | 24% |
| Domain 3 | Applications of Foundation Models (FMs) | 28% |
| Domain 4 | Guidelines for Responsible AI | 14% |
| Domain 5 | Security, Compliance, and Governance for AI Systems | 14% |
This is the order I used while studying:
Domain 1: Fundamentals of AI and Machine Learning
This domain covers ML fundamentals and model evaluation.
1. Machine Learning Paradigms and Core Algorithms
Machine learning workflows depend on how the data is labeled and what the model is trying to optimize.
- Supervised Learning: Data includes inputs and target labels.
- Classification: predicts categories such as spam vs. non-spam or disease diagnosis.
- Regression: predicts continuous values such as price, temperature, or forecast numbers.
- Unsupervised Learning: Data has no labels and the model discovers patterns.
- Clustering: groups similar data points together.
- Dimensionality Reduction: reduces the number of features while keeping the important structure.
- Reinforcement Learning: an agent learns by interacting with an environment and maximizing rewards.
2. Model Diagnostics: Overfitting, Underfitting, and Regularization
| Concept | Training Error | Validation / Test Error | Root Cause | Remediation |
|---|---|---|---|---|
| Underfitting | High | High | Model is too simple | Increase complexity, add features, reduce regularization |
| Overfitting | Low | High | Model memorizes noise | Add L1/L2 regularization, dropout, early stopping, more data |
| Optimal Fit | Low | Low | Balanced generalization | Maintain model capacity and validation performance |
3. Evaluation Metrics and Optimization Goals
| Metric | Definition | Business Focus |
|---|---|---|
| Precision | TP / (TP + FP) | Reduce false positives |
| Recall | TP / (TP + FN) | Reduce false negatives |
| F1-Score | Harmonic mean of precision and recall | Balanced metric for imbalanced data |
| RMSE / MAE | Regression error metrics | Understand prediction error magnitude |
Domain 1 Practice Questions
Question 1
A financial fraud investigation team is deploying a machine learning model to flag suspicious transactions. Missing a fraudulent transaction costs the company much more than manually reviewing a false alarm. Which metric should the team prioritize?
- A. Precision
- B. Recall
- C. Accuracy
- D. Specificity
Show answer and reason
Answer: B. Recall
Reason: Recall focuses on catching actual positives. When false negatives are expensive, recall matters more than raw accuracy.
Question 2
A neural network achieves 99% accuracy on the training data but only 62% on validation data. What should be done?
- A. Increase model capacity
- B. Remove regularization
- C. Apply L2 regularization and dropout
- D. Decrease the training dataset
Show answer and reason
Answer: C. Apply L2 regularization and dropout.
Reason: This is a classic overfitting pattern. Regularization helps the model generalize instead of memorizing training noise.
Question 3
An e-commerce company wants to segment customers without labels. Which approach is best?
- A. Logistic Regression
- B. K-Means Clustering
- C. Q-Learning
- D. Random Forest Regressor
Show answer and reason
Answer: B. K-Means Clustering.
Reason: This is an unsupervised learning task and clustering is the right pattern for grouping similar customer profiles.
Domain 2: Fundamentals of Generative AI
I spent the most time on this domain because it connects directly to how modern AI products work.
1. Transformer Architecture and Foundation Models
The Transformer architecture is the basis of most modern generative AI systems.
- Self-Attention: allows the model to weigh the importance of different tokens in the same sequence.
- Embeddings: convert tokens into dense vector representations.
- Decoder-only models: optimized for generation, such as GPT-style models.
- Encoder-only models: optimized for understanding and classification.
- Encoder-decoder models: used for sequence-to-sequence tasks like translation and summarization.
2. Inference Parameters and Output Control
| Parameter | Function | High Value Effect | Low Value Effect |
|---|---|---|---|
| Temperature | Controls randomness | More creative outputs | More deterministic outputs |
| Top-P | Uses cumulative probability threshold | More varied selection | More constrained output |
| Top-K | Limits candidate tokens | More diverse generation | More limited generation |
| Max Tokens / Stop Sequences | Controls response length | Prevents runaway output | Enforces stricter limits |
Domain 2 Practice Questions
Question 1
A team needs the exact same output format every time for contract extraction. Which setting is most appropriate?
- A. Temperature 1.0, Top-P 0.9
- B. Temperature 0.0, Top-P 0.1
- C. Top-K 500
- D. High presence penalty
Show answer and reason
Answer: B. Temperature 0.0 and Top-P 0.1.
Reason: Low randomness helps the model produce exact, repeatable outputs.
Question 2
Which mechanism in the Transformer allows the model to relate distant words in a sequence?
- A. Recurrent layers
- B. Self-Attention
- C. Convolutional kernels
- D. SGD
Show answer and reason
Answer: B. Self-Attention.
Reason: Self-attention enables parallel processing and captures context across long sequences.
Question 3
An organization wants numerical vector representations of text for semantic search. Which model type should it choose?
- A. Autoregressive decoder model
- B. Embedding model
- C. Diffusion model
- D. GAN
Show answer and reason
Answer: B. Embedding model.
Reason: Embedding models convert text into semantic vectors that work well for retrieval and search.
Domain 3: Applications of Foundation Models
This domain focuses on applying foundation models to real business problems, so I spent time comparing the main adaptation techniques.
1. Model Customization and Adaptation Spectrum
| Technique | Description | Use Case | Cost and Complexity |
|---|---|---|---|
| Prompt Engineering | Guides the model with instructions or examples | Quick tasks and prototypes | Lowest |
| RAG | Retrieves relevant information from a knowledge base before answering | Grounded answers, current docs | Low to moderate |
| Fine-Tuning | Updates a model with domain-specific examples | Tone, formatting, task adaptation | Moderate |
| Continued Pre-Training | Trains on large domain-specific corpora | Specialized domains | High |
2. AWS Native Generative AI Stack
- Amazon Bedrock: managed access to foundation models from multiple providers
- Knowledge Bases for Amazon Bedrock: managed retrieval and grounding workflow
- Agents for Amazon Bedrock: orchestrates multi-step workflows and tool use
- Guardrails for Amazon Bedrock: applies content filtering, PII handling, and topic restrictions
- Amazon Q: AI assistant for work and developer productivity
- SageMaker JumpStart: access to pretrained models and solution templates
Domain 3 Practice Questions
Question 1
A healthcare company needs answers based on internal documentation updated daily, with source citations and low hallucination risk. What is the best fit?
- A. Continued pre-training
- B. LoRA fine-tuning
- C. Knowledge Bases for Amazon Bedrock with RAG
- D. Static few-shot prompts
Show answer and reason
Answer: C. Knowledge Bases for Amazon Bedrock with RAG.
Reason: RAG surfaces relevant documents at inference time and lets the system cite the source documents.
Question 2
A financial app needs an agent to check account balances, convert currencies, and execute transfers. Which AWS capability supports this workflow?
- A. Model customization
- B. Agents for Amazon Bedrock
- C. Provisioned throughput
- D. Multimodal embeddings
Show answer and reason
Answer: B. Agents for Amazon Bedrock.
Reason: Bedrock Agents break a task into steps and call the right APIs or tools.
Question 3
An engineering team wants to compare two summarization models using human reference summaries. Which evaluation metrics should they use?
- A. Precision and Recall
- B. ROUGE and BLEU
- C. Log-loss and Gini impurity
- D. MAE and R-squared
Show answer and reason
Answer: B. ROUGE and BLEU.
Reason: These are standard metrics for text generation and summarization quality.
Domain 4: Guidelines for Responsible AI
This domain focuses on making AI systems trustworthy, fair, and understandable.
Core Pillars
- Fairness: reduce demographic or historical bias
- Explainability and Interpretability: explain why a model made a decision
- Robustness and Safety: resist attacks and unexpected edge cases
- Privacy and Transparency: protect sensitive data and communicate usage clearly
AWS Tools for Responsible AI
| AWS Tool / Service | Capability |
|---|---|
| Amazon SageMaker Clarify | Bias detection and feature attribution explanations |
| Amazon Bedrock Guardrails | Safety policies, PII redaction, toxic content blocking |
| AWS AI Service Cards | Documentation on intended use and safety boundaries |
| Amazon A2I | Human review workflow for low-confidence predictions |
Domain 4 Practice Questions
Question 1
A loan model rejects applicants from one demographic at a much higher rate. Compliance wants to explain individual decisions. Which AWS tool should be used?
- A. CloudWatch Logs
- B. SageMaker Clarify
- C. CloudTrail
- D. Bedrock Provisioned Throughput
Show answer and reason
Answer: B. SageMaker Clarify.
Reason: Clarify helps with bias detection and feature attribution for decision explanations.
Question 2
A chatbot must block harmful requests, filter financial advice, and redact social security numbers. What is the best AWS feature?
- A. SageMaker Model Monitor
- B. Bedrock Guardrails
- C. AWS Shield Advanced
- D. IAM policies
Show answer and reason
Answer: B. Bedrock Guardrails.
Reason: Guardrails can apply safety policies, denied topics, and PII redaction.
Question 3
An insurance company routes low-confidence claims predictions to a human reviewer. Which AWS service supports this?
- A. Amazon A2I
- B. AWS Step Functions
- C. SageMaker Ground Truth
- D. CodePipeline
Show answer and reason
Answer: A. Amazon A2I.
Reason: A2I is designed for human-in-the-loop review of ML predictions.
Domain 5: Security, Compliance, and Governance for AI Systems
This domain connects model quality to secure data handling, governance, and ongoing monitoring.
1. Data Protection and Security Controls in Amazon Bedrock
- Shared responsibility model: AWS secures the infrastructure, while customers secure their data, prompts, outputs, and IAM control.
- Data privacy: prompts and fine-tuning data are not used to train base AWS foundation models.
- Encryption: TLS protects data in transit, and KMS customer-managed keys protect data at rest.
- Network isolation: use VPC endpoints and PrivateLink to keep traffic off the public internet.
2. Monitoring, Auditing, and Compliance Tools
| Service | Function |
|---|---|
| AWS CloudTrail | Logs API calls for governance and audit |
| Amazon CloudWatch | Tracks latency, invocation counts, and errors |
| Amazon SageMaker Model Monitor | Detects data drift and concept drift |
| AWS Artifact | Provides compliance reports and security attestations |
Domain 5 Practice Questions
Question 1
A security team wants private connectivity to Bedrock without public internet exposure. Which combination is correct?
- A. Client VPN and Internet Gateway
- B. PrivateLink VPC Endpoints and Amazon VPC
- C. WAF and CloudFront
- D. Direct Connect with public virtual interfaces
Show answer and reason
Answer: B. PrivateLink VPC Endpoints and Amazon VPC.
Reason: PrivateLink gives private path connectivity from inside the VPC to Bedrock endpoints.
Question 2
An organization asks whether prompts sent to Bedrock are used to improve foundation models. What is the default policy?
- A. Data is stored for 30 days and used for retraining
- B. Prompts and responses are private to the account and not used to train base models
- C. Fine-tuning data is publicly visible in the region
- D. Data is shared with model providers through a non-disclosure agreement
Show answer and reason
Answer: B. Prompts and responses are private to the account and not used to train base models.
Reason: This is one of the key Bedrock privacy guarantees.
Question 3
Production model accuracy drops over time because the input distribution shifts. What is this, and what service detects it?
- A. Overfitting; CloudTrail
- B. Data drift; SageMaker Model Monitor
- C. Hyperparameter misconfiguration; AWS Config
- D. Adversarial poisoning; Guardrails
Show answer and reason
Answer: B. Data drift; SageMaker Model Monitor.
Reason: Model monitoring is specifically used to detect drift in production ML systems.
Practice Scenario Questions
Scenario 1: Fraud Detection
A financial services company wants to detect suspicious transactions in real time. Missing a true fraud event is a much bigger business problem than investigating a false alarm. Which metric should guide the model selection process?
- A. Precision
- B. Recall
- C. Accuracy
- D. Specificity
Show answer and reason
Answer: B. Recall
Explanation: Recall focuses on identifying fraud cases correctly. The cost of missing a real fraud case is higher than the cost of a manual review.
Scenario 2: Legal Document Extraction
A legal team needs a system to extract standard clauses from contracts and return the exact text format every time. Which inference settings would be best?
- A. Temperature 1.0 and Top-P 0.9
- B. Temperature 0.0 and Top-P 0.1
- C. High Top-K and high temperature
- D. High frequency penalty
Show answer and reason
Answer: B. Temperature 0.0 and Top-P 0.1.
Explanation: This reduces randomness and makes the output deterministic.
Scenario 3: Support Bot with Internal Documentation
A company wants a chatbot to answer questions using internal policy documents and cite the source pages. Which solution is best?
- A. Continued pre-training
- B. LoRA fine-tuning
- C. Bedrock Knowledge Bases with RAG
- D. Static prompt injection only
Show answer and reason
Answer: C. Bedrock Knowledge Bases with RAG.
Explanation: RAG allows the model to use current internal knowledge and provide citations.
Scenario 4: Responsible AI Review
A hiring workflow is rejecting candidates from a particular group at a much higher rate. Compliance wants feature-level explanations for a decision. Which service should be used?
- A. CloudWatch
- B. SageMaker Clarify
- C. CloudTrail
- D. AWS WAF
Show answer and reason
Answer: B. SageMaker Clarify.
Explanation: Clarify helps detect bias and provides explainability through feature attribution.
Scenario 5: Bedrock Security
A company wants to keep prompts and outputs private and prevent them from being used to train public foundation models. What is the correct AWS stance?
- A. Data is stored in the public model registry
- B. Data is kept private to the AWS account and not used to train base models
- C. Data is used for model improvement unless the user opts out
- D. Data is stored in a public bucket by default
Show answer and reason
Answer: B. Data is kept private to the AWS account and not used to train base models.
Explanation: Bedrock does not use customer inputs for base model training.
Scenario 6: Production Monitoring
A production image classification model slowly becomes less accurate because the underlying data distribution has changed. What is happening?
- A. Overfitting
- B. Data drift
- C. Prompt injection
- D. Token leakage
Show answer and reason
Answer: B. Data drift.
Explanation: The real-world input distribution has shifted relative to the original training data.
Five Practice Questions
These are additional review questions covering the five exam domains. They are not intended to represent a complete mock exam.
Practice Question 1: ML Fundamentals
A data science team is preparing a tabular dataset for a linear regression model. Several numerical features have very different ranges, such as age and annual income. Which step should be applied?
- A. One-hot encoding
- B. Feature scaling and normalization
- C. Target labeling
- D. PCA
Show answer and reason
Answer: B. Feature scaling and normalization.
Explanation: Scaling keeps features on a comparable range and prevents larger-magnitude features from dominating training.
Practice Question 2: Generative AI Fundamentals
A company wants to prevent a model from generating endless or excessive output. Which parameter sets the hard ceiling?
- A. Temperature
- B. Max Tokens
- C. Top-P
- D. Frequency Penalty
Show answer and reason
Answer: B. Max Tokens.
Explanation: Max tokens sets the maximum response length and prevents runaway generation.
Practice Question 3: Applications of Foundation Models
A company wants to build a customer support chatbot that uses internal documentation and provides source citations. Which solution fits best with the least operational effort?
- A. Fine-tune a Titan model on internal docs
- B. Train a custom model from scratch on SageMaker HyperPod
- C. Use Knowledge Bases for Amazon Bedrock connected to OpenSearch Serverless
- D. Build a custom neural network on Deep Learning AMIs
Show answer and reason
Answer: C. Use Knowledge Bases for Amazon Bedrock connected to OpenSearch Serverless.
Explanation: This is a managed RAG workflow that retrieves relevant records and includes citations.
Practice Question 4: Responsible AI
An HR team uses an AI tool to screen candidates. Auditors want clear explanations for why a candidate was rejected. Which pillar does this address, and which AWS service helps?
- A. Sustainability with AWS Carbon Footprint Tool
- B. Explainability with SageMaker Clarify
- C. Security with AWS KMS
- D. Privacy with Amazon Macie
Show answer and reason
Answer: B. Explainability with SageMaker Clarify.
Explanation: Explainability is about understanding the reasons behind a model decision.
Practice Question 5: Security and Governance
An enterprise wants to encrypt Bedrock knowledge base data with keys they directly manage. Which service configuration meets that requirement?
- A. AWS KMS with Customer Managed Keys
- B. Secret keys in environment variables
- C. Default AWS managed keys in S3
- D. IAM inline policies
Show answer and reason
Answer: A. AWS KMS with Customer Managed Keys.
Explanation: Customer Managed Keys give the organization direct control over the encryption lifecycle and policies.
Final Thoughts
The biggest lesson for me was that AIF-C01 is not just about memorizing model names or AWS service labels. Connecting each concept to a real AWS service and business use case made the material easier to understand and remember, and helped me see how AI systems are designed, evaluated, and governed in production.
My advice