AIF-C01 study cheatsheet

One-line cues for terms and in-scope AWS services. Unofficial; verify against the current exam guide.

Not from AWS. Use alongside aif-c01-in-scope-services.html and your domain decks. Lists and product names change over time.

In-scope services (slide deck) · Official in-scope list

Terms Patterns Responsible AI Security Services

Core terms

AI / ML / GenAI
Artificial intelligence (AI)	Systems that perform tasks requiring human-like reasoning, perception, or decision support.
Machine learning (ML)	Learns patterns from data to predict or decide without explicit step-by-step programming for every case.
Deep learning	ML using layered neural nets—strong for images, text, speech at scale when data/compute allow.
Supervised learning	Training on labeled input→output pairs (classification, regression).
Unsupervised learning	Find structure without labels (clustering, anomaly patterns, dimensionality reduction).
Reinforcement learning	Agent learns via rewards/penalties from environment interactions over time.
Training	Fitting model parameters on historical data; typically batch, compute-heavy, offline.
Inference	Running a trained model on new inputs for predictions or generated output.
Fine-tuning	Adapting a pretrained model to a narrower domain/task with extra labeled or instruction data.
Foundation model (FM)	Large pretrained model (often transformers) generalizing across many downstream tasks via prompting or adaptation.
Generative AI	Models that synthesize new content (text, code, images) vs only classifying or scoring inputs.
Token	Chunk of text a model processes/bills; context length = max tokens the model can consider at once.
Embedding	Dense vector representing text/image meaning for similarity search and retrieval.
Hallucination	Confident but false or unsupported model output; mitigated by grounding, guardrails, and evaluation.
Prompt	Instruction + context given to an FM; prompt engineering shapes quality without weight updates.
Temperature	Sampling randomness—lower = more deterministic; higher = more creative/varied.
RAG	Retrieve relevant documents/chunks, then condition the FM—reduces hallucination and adds freshness.
Vector store / k-NN	Stores embeddings and returns nearest neighbors for semantic search in RAG pipelines.
Agent	FM + tools (API, search, code) that plans steps or calls actions, not single-shot completion.
MLOps	Operational practices for reproducible training, deployment, monitoring, and rollback of ML.
Drift	Data or concept change over time degrading model quality; needs monitoring and retraining.
Bias (ML)	Systematic skew in predictions across groups—address with data, evaluation, and constraints.
Explainability	Evidence of why a model produced an output (SHAP, attention, counterfactuals, simplified models).

Metrics (recognition)
Accuracy	Fraction of correct predictions across all classes.
Precision	Of positive predictions, how many were true positives—care about false alarms in the positive class.
Recall	Of actual positives, how many were found—care about missing positives.
F1	Harmonic balance of precision and recall when you need one scalar score.
Confusion matrix	Counts of TP/FP/TN/FN—basis for precision/recall and class imbalance analysis.

Architecture patterns (exam stems)

“Which AWS …?” shortcuts
Managed FM API	Bedrock (models, Knowledge Bases, Guardrails, Agents).
Custom train + deploy lifecycle	SageMaker AI; quick starts: JumpStart.
Enterprise doc Q&A with connectors	Kendra vs self-built RAG on OpenSearch/pgvector—match stem to “managed search product” vs DIY vectors.
Vector similarity at scale	OpenSearch k-NN; Aurora/RDS Postgres extensions; Neptune for graph+vector cases.
Speech in / text out	Transcribe.
Text in / audio out	Polly.
OCR / forms from PDF	Textract.
Image labels / moderation	Rekognition.
Chatbot with intents	Lex vs open-ended FM chat—Lex for structured dialog.
Recommendations	Personalize.
Human in the loop	A2I for review when confidence low or policy requires humans.
Datasets & artifacts	S3; long-term retention: Glacier.

Responsible AI (pillars → tools)

Fairness, explainability, privacy, robustness
Fairness	Measure disparate impact; Clarify (bias metrics), balanced data, Guardrails content/topic filters.
Explainability	Clarify explainability; SHAP-style insights; human-readable rationales where feasible.
Privacy	Minimize PII in prompts/training; Macie on S3; KMS encryption; avoid logging secrets.
Robustness / safety	Guardrails (Bedrock), content filters, red-teaming, evaluation harnesses.
Transparency	Document data lineage, model cards, human review (A2I) for high-stakes decisions.

Security & governance (shared + AWS)

Controls that appear in scenario questions
Least privilege	IAM roles scoped to invoke model, read S3 prefix, use KMS key.
Audit who did what	CloudTrail API history.
Encryption keys	KMS CMKs for data at rest; TLS in transit.
Secrets in apps	Secrets Manager (rotation) vs hard-coding API keys.
Config drift	AWS Config rules on encryption, public access, tags.
Compliance evidence	Audit Manager; Artifact for reports.

In-scope services (exam one-liners)

Analytics
AWS Data Exchange	Subscribe to third-party datasets inside AWS for ML/analytics.
Amazon EMR	Distributed Spark/Hadoop for large-scale data prep before ML.
AWS Glue	Serverless ETL, crawlers, Data Catalog for lake→analytics/ML pipelines.
AWS Glue DataBrew	Visual data cleaning/normalization for analysts.
AWS Lake Formation	Governance and fine-grained access on data lake assets.
Amazon OpenSearch Service	Search + k-NN/vectors; common RAG retrieval backend.
Amazon QuickSight	BI dashboards and metrics—not model training.
Amazon Redshift	Cloud data warehouse; SQL analytics and joins for ML features.

Cloud financial management
AWS Budgets	Alerts and limits when spend exceeds plans (incl. AI workloads).
AWS Cost Explorer	Slice and analyze historical AWS costs by dimension.

Compute & containers
Amazon EC2	VMs/GPUs for custom training or BYO hosting when managed APIs are not enough.
AWS Lambda	Event-driven glue: invoke Bedrock/SageMaker, lightweight orchestration.
Amazon ECS	Run containerized inference/training workers with AWS-native orchestration.
Amazon EKS	Managed Kubernetes for portable, complex ML microservices.

Database
Amazon Aurora	Relational; often Postgres-compatible patterns incl. vectors with extensions.
Amazon RDS	Managed SQL; vector-on-Postgres option for RAG indexes.
Amazon DocumentDB	Mongo-compatible documents; document + embedding scenarios.
Amazon DynamoDB	Low-latency key-value for app state, metadata—not primary semantic search engine.
Amazon ElastiCache	In-memory cache to cut latency or repeated retrieval cost.
Amazon Neptune	Graph DB; relationship-heavy knowledge + graph/vector patterns.

Developer tools
Kiro	AWS builder IDE/agent positioning—recognition for “accelerate development with AI.”
Strands Agents	Agent builder tooling on AWS—recognition-level.
Amazon Q	GenAI assistant in IDE, console, and business workflows.

Machine learning — A–L
Amazon A2I	Human review workflows for low-confidence or regulated predictions.
Amazon Bedrock	Primary managed FM layer: models, Knowledge Bases, Guardrails, Agents.
Amazon Bedrock AgentCore	Building blocks for production agents on Bedrock.
Amazon Comprehend	Managed NLP: entities, sentiment, classification, PII.
Amazon Kendra	Enterprise intelligent search with connectors—contrast with DIY vector on OpenSearch.
Amazon Lex	Conversational bots (intents/slots) for chat and voice.
Amazon Nova	AWS FM family in Bedrock ecosystem—FM choice recognition.

Machine learning — M–Z
Amazon Personalize	Real-time recommendations for retail/media use cases.
Amazon Polly	Text-to-speech synthesis.
Amazon Rekognition	Image/video analysis: labels, moderation, faces.
Amazon SageMaker AI	End-to-end custom ML build/train/deploy/monitor.
SageMaker JumpStart	Prebuilt models and solution templates in SageMaker.
Amazon Textract	Extract printed/handwritten text, forms, tables from documents.
Amazon Transcribe	Speech-to-text for audio/video.
Amazon Translate	Neural machine translation.
AWS Transform	Modernization toward cloud-ready analytics/AI estates (per guide scope—verify naming).

Management & governance
AWS CloudTrail	API audit log for governance and investigations.
Amazon CloudWatch	Metrics, alarms, logs for operational health.
AWS Config	Resource inventory and rule-based compliance checks.
AWS Trusted Advisor	Automated best-practice checks (cost, security, limits).
AWS Well-Architected Tool	Structured workload reviews vs the Well-Architected pillars.

Networking & CDN
Amazon VPC	Private networks, subnets, SGs—foundation for private AI endpoints and data paths.
Amazon CloudFront	Global CDN for low-latency delivery of apps and static assets.

Security, identity, compliance
AWS IAM	Users, roles, policies controlling access to models and data.
AWS KMS	Encryption keys for data and secrets at rest.
AWS Secrets Manager	Store and rotate application secrets (API keys, DB creds).
Amazon Macie	Discover sensitive data (e.g. PII) in S3 before training or indexing.
Amazon Inspector	Vulnerability scanning for EC2/ECR workloads (recognition).
AWS Artifact	Download compliance reports and agreements.
AWS Audit Manager	Continuous control evidence mapped to frameworks.

Storage
Amazon S3	Object store for datasets, features, model artifacts, logs.
Amazon S3 Glacier	Low-cost archival for long-retention model or data history.

End of cheatsheet. Print this page from your browser if you want a PDF.