AWS Certified AI Practitioner (AIF-C01) Master Exam Cheat Sheet
1. Master Scenario Anchor Cheat Sheet ("If You See X, Choose Y")
| If the Exam Scenario Mentions... | Choose This AWS Service / Solution | Why This Solution Wins |
|---|---|---|
| "Extract text, tables, and handwriting from scanned PDFs without custom ML" | Amazon Textract | Pre-trained, fully managed document extraction API. |
| "Detect faces, custom objects, or moderate inappropriate images/video" | Amazon Rekognition | Pre-trained computer vision API with zero model training required. |
| "Build conversational voice or text chatbots using natural language understanding" | Amazon Lex | Managed conversational AI engine powering automated bots. |
| "Convert written text into lifelike spoken audio across multiple languages" | Amazon Polly | Pre-trained text-to-speech (TTS) neural audio API. |
| "Transcribe audio call recordings into time-stamped text files" | Amazon Transcribe | Managed automatic speech recognition (ASR) service. |
| "Extract sentiment, entities, and key topics from unstructured customer text" | Amazon Comprehend | Pre-trained natural language processing (NLP) service. |
| "Deliver real-time personalized product recommendations based on user history" | Amazon Personalize | Pre-built recommendation engine powered by AWS ML. |
| "Incorporate dynamic, up-to-date company policies without retraining the model" | Amazon Bedrock Knowledge Bases (RAG) | Ground answers in external S3 vector stores to eliminate hallucinations. |
| "Adapt model output style, tone, persona, or custom output format (e.g., JSON)" | Supervised Fine-Tuning (SFT) | Updates internal model weights to enforce target behavior and formatting. |
| "Teach a model specialized domain technical jargon/acronyms using unlabeled text" | Continuous Pre-Training | Ingests unlabeled text to adapt base weights to specialized vocabulary. |
| "Filter toxic prompts, block PII, reject denied topics, and block hallucinations in real time" | Guardrails for Amazon Bedrock | Real-time safety filter applied directly to Bedrock inference requests. |
| "Detect demographic bias in datasets/models and calculate SHAP feature importance" | Amazon SageMaker Clarify | ML explainability and bias analysis engine for training and inference. |
| "Route low-confidence predictions or sensitive AI completions to human reviewers" | Amazon Augmented AI (Amazon A2I) | Managed human-in-the-loop workflow orchestrator for ML predictions. |
| "Discover, classify, and redact sensitive PII in S3 buckets before data ingestion" | Amazon Macie | Automated S3 security service using ML to detect PII and credentials. |
| "Route Bedrock API traffic privately inside VPC without traversing public internet" | AWS PrivateLink | Establishes private network endpoints between customer VPC and Bedrock. |
| "Enforce strict, declarative tool execution limits on agents independently of LLM reasoning" | Policy in Amazon Bedrock AgentCore | Gateway-level authorization controls that block unauthorized API calls. |
| "Manage workload identities, inbound auth, and outbound OAuth keys for AI agents" | Amazon Bedrock AgentCore Identity | Secure identity and token management runtime for autonomous agents. |
| "Download AWS SOC/ISO compliance reports or sign HIPAA BAA" | AWS Artifact | Central portal for downloading AWS compliance reports and legal agreements. |
| "Continuously monitor AWS resource configurations for compliance drift" | AWS Config | Evaluates resource states against security baselines and auto-remediates drift. |
| "Document custom model training data, architecture, performance, and governance" | Amazon SageMaker Model Cards | Customizable governance artifact for tracking custom ML models. |
| "AWS-provided non-customizable documentation for pre-trained AWS AI services" | AWS AI Service Cards | AWS-published documentation outlining capabilities and limits of pre-trained AI services. |
2. Domain 1 & 2: Core AI/ML & Generative AI Fundamentals
Learning Paradigms Quick Comparison
- Supervised Learning: Model trains on labeled data ($X \rightarrow Y$).
- Classification: Predicts discrete categorical labels (e.g., spam/not spam, churn/no churn).
- Regression: Predicts continuous numerical values (e.g., house prices, temperature).
- Unsupervised Learning: Model trains on unlabeled data to find hidden patterns.
- Clustering: Groups similar data points without prior class labels (e.g., customer segmentation).
- Reinforcement Learning: Agent interacts with an environment, learning optimal policies via rewards and penalties.
Model Fitting & Parameter Adjustments
$$\text{High Bias} = \text{Underfitting (Model too simple)} \quad \Big\vert{} \quad \text{High Variance} = \text{Overfitting (Model memorizes noise)}$$
- Overfitting Mitigation: Increase training data, apply regularization, simplify model architecture, or reduce training epochs.
- Underfitting Mitigation: Increase model complexity, add relevant features, reduce regularization, or train for more epochs.
- Hyperparameters vs. Weights:
- Hyperparameters: External settings configured before training (e.g., learning rate, batch size, epochs).
- Weights & Biases: Internal parameters learned and updated during training.
SageMaker Inference Options Selection Matrix
| SageMaker Inference Type | Operational Characteristics | Ideal Scenario |
|---|---|---|
| Real-Time Inference | Sub-second latency, persistent dedicated instances. | Sustained high-volume interactive web apps. |
| Serverless Inference | Scales down to zero when idle; pay-per-execution. | Unpredictable, intermittent traffic with long idle gaps. |
| Asynchronous Inference | Payloads up to 1 GB, processing time up to 1 hour, built-in queue. | Large file processing (e.g., high-res video) with queuing. |
| Batch Transform | Offline batch execution; resources terminate on completion. | Scheduled overnight bulk dataset scoring (e.g., daily risk reports). |
3. Domain 3: Foundation Model Applications & Design
The FM Customization Spectrum
```text [Prompt Engineering] ---> [RAG] ---> [Distillation] ---> [Fine-Tuning] ---> [Pre-Training] (Lowest Cost/Effort) (Highest Cost/Effort)