AWS Certified AI Practitioner 3-Week Study Plan

Applications of Foundation Models

~18 Questions. Amazon Bedrock foundation models (Claude, Titan, Llama), Knowledge Bases for RAG, Bedrock Agents, Guardrails, SageMaker JumpStart, and Amazon Q enterprise tools.

Responsible AI, Governance & Security

~18 Questions. Pillars of responsible AI, bias detection with SageMaker Clarify, explainability (SHAP), data drift monitoring, Shared Responsibility Model, and Bedrock privacy guarantees.

Fundamentals of Generative AI

~16 Questions. Transformers, self-attention, tokenization, context windows, inference parameters (Temperature, Top-P, Top-K), prompt engineering (zero/few-shot, CoT), vector embeddings, and RAG vs fine-tuning.

Fundamentals of AI and ML

~13 Questions. AI vs ML vs Deep Learning taxonomy, Supervised (regression/classification), Unsupervised (clustering), Reinforcement Learning, overfitting/underfitting, and evaluation metrics (Precision, Recall, F1).

• Master the concentric hierarchy: Deep Learning ⊂ Machine Learning ⊂ Artificial Intelligence.
• Differentiate learning paradigms: Supervised (regression for continuous, classification for discrete), Unsupervised (K-Means clustering), and Reinforcement Learning (agent, environment, reward).
• Understand the bias-variance tradeoff: Underfitting (high bias, fix with model complexity) vs. Overfitting (high variance, fix with regularization, dropout, and data augmentation).
• Calculate classification metrics: Precision (TP / (TP + FP)), Recall (TP / (TP + FN)), and F1 Score (harmonic mean for imbalanced data).
• Study Generative AI fundamentals: Transformer self-attention parallelization, subword tokenization (~750 words per 1,000 tokens), and context window limits.
• Tune inference parameters: Low Temperature (0.0–0.2) for deterministic compliance vs. high Temperature for creativity; Top-P (nucleus cumulative probability sampling) and Top-K.
• Practice prompt engineering patterns: Zero-shot, Few-shot (in-context exemplars), and Chain-of-Thought (step-by-step reasoning).

• Deep dive into Amazon Bedrock: Serverless API consumption of top foundation models without infrastructure management.
• Navigate the Bedrock model catalog: Anthropic Claude 3.5 Sonnet (coding and complex reasoning), Amazon Titan (cost-effective text and embeddings), and Meta Llama.
• Master Knowledge Bases for Amazon Bedrock: Automated end-to-end RAG pipelines syncing S3 documents, chunking with overlap, generating vector embeddings, and indexing into OpenSearch Serverless.
• Understand Agents for Amazon Bedrock: Orchestrating multi-step tasks and calling internal APIs via AWS Lambda Action Groups backed by OpenAPI schemas.
• Configure Guardrails for Amazon Bedrock: Implementing Denied Topics, Content Filters, Sensitive Information (PII) masking, and Contextual Grounding Checks to detect hallucinations.
• Compare Bedrock vs. SageMaker JumpStart: Serverless API vs. dedicated instance hosting with full VPC and container control.
• Explore Amazon Q assistants: Amazon Q Developer (IDE coding, testing, Java transformations) vs. Amazon Q Business (enterprise connector assistant with strict IAM ACL sync).

• Master the 6 pillars of Responsible AI: Fairness, Explainability, Privacy, Robustness, Transparency, and Governance.
• Audit bias using Amazon SageMaker Clarify: Pre-training demographic disparity and post-training disparate impact metrics.
• Generate explainability using SHAP values: Quantifying feature attribution for adverse action compliance.
• Monitor live production models using Amazon SageMaker Model Monitor: Detecting Data Drift and Concept Drift.
• Apply the AWS Shared Responsibility Model for AI: Customer owns data, IAM permissions, prompt safety, and output validation; AWS owns infrastructure, model weights, and physical security.
• Memorize the Bedrock Data Privacy Promise: Customer prompts and outputs are NEVER used to train base foundation models and NEVER shared with model providers.
• Review security architecture: AWS KMS Customer Managed Keys (CMKs) for encryption at rest, TLS 1.3 in transit, AWS PrivateLink Interface VPC Endpoints, and AWS CloudTrail auditing.
• Take full 65-question timed mock exams on our interactive simulator and conduct distractor autopsies on every missed item.