ExamFinch.com My Study

Cheat sheet

AWS AI Practitioner Cheat Sheet

Practice this exam

AI/ML Foundations

  • 20% of exam
  • AI Terms Learning Types Use Cases Lifecycle Metrics

GenAI Foundations

  • 24% of exam
  • Foundation Models Tokens Embeddings Prompts Inference

Foundation Model Apps

  • 28% of exam
  • Amazon Bedrock RAG Agents Guardrails Service Picker

Responsible AI

  • 14% of exam
  • Fairness Explainability Transparency Human Review Safety

Security + Governance

  • 14% of exam
  • IAM Encryption Logging Compliance Cost Controls

Quick Facts

  • Exam AIF-C01
  • Level Foundational
  • Questions 65 total
  • Scored 50
  • Time 90 min
  • Pass 700/1000
  • Cost $100 USD
  • Validity 3 years

Supervised vs Unsupervised FO

Supervised

  • Labeled data
  • Known answers
  • Classification/regression

Unsupervised

  • Unlabeled data
  • Find patterns
  • Clustering/anomaly
  • Labels vs patterns

Use-Case Picker FO

  • Known labels → Supervised learning
  • Hidden groups → Clustering
  • Numeric outcome → Regression
  • Future time series → Forecasting
  • Next-best item → Recommendation
  • Policy rewards → Reinforcement learning

AI Terms FO

  • AI Human-like tasks
  • ML Learns from data
  • Deep learning Neural-network ML
  • Model Learned pattern
  • Training Learn parameters
  • Inference Generate prediction
  • Feature Input signal
  • Label Known answer

Batch vs Real-Time FO

Batch

  • Many records
  • Offline timing
  • Lower urgency

Real-time

  • Single request
  • Immediate response
  • Latency critical
  • Bulk vs instant

Learning Types FO

  • Supervised Labeled data
  • Unsupervised Unlabeled patterns
  • Reinforcement Reward feedback
  • Classification Predict class
  • Regression Predict number
  • Clustering Group similar records
  • Forecasting Predict future values
  • Recommendation Suggest next item

ML Lifecycle FO

  • Collect Gather data
  • Prepare Clean data
  • Train Fit model
  • Evaluate Measure quality
  • Deploy Serve model
  • Monitor Watch drift
  • Retrain Refresh model
  • MLOps Repeatable ML operations

GenAI Stack GE

  • Prompt, retrieve, guard, evaluate
  • Prompt directs RAG grounds Guardrails constrain Eval proves

RAG vs Fine-Tune GE

RAG

  • Adds knowledge
  • Uses retrieval
  • Keeps model

Fine-tune

  • Changes behavior
  • Needs training data
  • Updates model
  • Knowledge vs behavior

Model Pattern Picker GE

  • General task → Prompting
  • Private knowledge → RAG
  • Style/behavior shift → Fine-tuning
  • External actions → Agent
  • Safety constraints → Guardrails
  • Quality proof → Evaluation

GenAI Terms GE

  • FM Broad pretrained model
  • LLM Text foundation model
  • Token Text unit
  • Context window Prompt memory size
  • Embedding Vector representation
  • Vector store Embedding search index
  • Hallucination Unsupported output
  • Grounding Answer from sources

Temperature vs Top-p GE

Temperature

  • Randomness level
  • Higher creative
  • Lower deterministic

Top-p

  • Token pool
  • Probability cutoff
  • Nucleus sampling
  • Randomness vs pool

Prompting GE

  • Zero-shot No examples
  • Few-shot Examples included
  • Template Reusable prompt
  • Role prompt Persona/instructions
  • Temperature Randomness control
  • Top-p Token pool control
  • Stop sequence Output cutoff
  • Prompt injection Instruction attack

Domain Weights AP

  • 20, 24, 28, 14, 14
  • Apps largest GenAI second Governance equals responsible

Bedrock vs SageMaker AP

Bedrock

  • Managed FMs
  • RAG/agents
  • No infrastructure

SageMaker AI

  • Custom ML
  • Train/deploy
  • Full lifecycle
  • FM apps vs ML platform

Service Picker AP

  • Managed FMs → Bedrock
  • Custom ML lifecycle → SageMaker AI
  • Text sentiment → Comprehend
  • Extract documents → Textract
  • Image labels → Rekognition
  • Speech transcript → Transcribe

Amazon Bedrock AP

  • Bedrock Managed FM platform
  • Model access Enable FMs
  • Knowledge Bases Managed RAG
  • Prompt Management Versioned prompts
  • AgentCore Build and run agents
  • Guardrails Safety controls
  • Evaluation Model quality checks
  • Provisioned throughput Reserved capacity

Service Memory AP

  • Text, image, document, speech
  • Comprehend text Rekognition image Textract document Transcribe speech

FM Patterns AP

  • Prompting Instruction only
  • RAG Retrieve context
  • Fine-tuning Adapt model behavior
  • Pretraining Build base model
  • Agent Plans tool use
  • Embedding search Semantic retrieval
  • Human review Manual judgment
  • Red team Adversarial testing

AWS Services AP

  • SageMaker AI ML build/deploy
  • Comprehend NLP insights
  • Rekognition Image/video analysis
  • Textract Document extraction
  • Transcribe Speech to text
  • Translate Language translation
  • Lex Chatbot voice/text
  • Polly Text to speech

Responsible AI RE

  • Fairness Reduce bias
  • Explainability Understand reasoning
  • Transparency Disclose AI use
  • Privacy Protect personal data
  • Safety Limit harmful output
  • Governance Policies and ownership
  • Bias Skewed outcomes
  • Human review People verify outputs

Security Memory GO

  • IAM, KMS, logs, budgets
  • IAM access KMS keys CloudTrail audit Budgets cost

Guardrails vs IAM GO

Guardrails

  • Output safety
  • Denied topics
  • Content filters

IAM

  • Access control
  • Principal permissions
  • AWS actions
  • Content vs access

Governance Picker GO

  • Restrict access → IAM
  • Encrypt data → KMS
  • Audit API calls → CloudTrail
  • Monitor metrics → CloudWatch
  • Find sensitive data → Macie
  • Control spend → Budgets

Security + Cost GO

  • IAM Least privilege
  • KMS Key management
  • CloudTrail API activity logs
  • CloudWatch Metrics and alarms
  • Config Resource compliance
  • Macie Sensitive data discovery
  • Budgets Cost alerts
  • Cost Explorer Spend analysis

Common Traps

Build vs use

  • AIF is practitioner ≠ Coding out-of-scope

RAG vs training

  • RAG retrieves ≠ Training changes model

Bedrock vs SageMaker

  • Bedrock runs FMs ≠ SageMaker builds ML

Guardrails vs security

  • Guardrails shape content ≠ IAM controls access

Accuracy vs fairness

  • Accuracy measures correctness ≠ Fairness checks bias

CloudTrail vs CloudWatch

  • CloudTrail logs API ≠ CloudWatch monitors metrics

Last Minute

  • 1. Apps domain largest: 28%
  • 2. 50 scored, 15 unscored
  • 3. Pass = 700 scaled
  • 4. Bedrock = managed FMs
  • 5. SageMaker AI = ML lifecycle
  • 6. RAG grounds with retrieval
  • 7. Fine-tune changes behavior
  • 8. Guardrails constrain content
  • 9. IAM controls AWS access
  • 10. Budgets alert on spend