AWS Certified AI Practitioner

1.1 Core AI & Machine Learning Concepts

• Defining AI, ML, Deep Learning, and their relationships — the AI/ML/DL hierarchy
• Types of machine learning: Supervised, Unsupervised, Semi-supervised, and Reinforcement Learning
• Key ML concepts: Training, Inference, Overfitting, Underfitting, Bias-Variance Tradeoff
• Neural networks: Perceptrons, activation functions, layers, forward propagation and backpropagation
• Natural Language Processing (NLP): tokenisation, stemming, embeddings, and sentiment analysis
• Computer Vision: image classification, object detection, and facial recognition concepts
• Anomaly detection, time-series forecasting, and recommendation systems in business contexts

1.2 ML Development Lifecycle on AWS

• The ML lifecycle: Problem Framing → Data Collection → Feature Engineering → Model Training → Evaluation → Deployment → Monitoring
• Data preparation: data labelling with Amazon SageMaker Ground Truth, feature stores, and data pipelines
• Model training approaches: SageMaker built-in algorithms, custom training scripts, and AutoML via SageMaker Autopilot
• Model evaluation metrics: Accuracy, Precision, Recall, F1-Score, AUC-ROC for classification; RMSE, MAE for regression
• Model deployment options: real-time endpoints, batch transform, serverless inference, and multi-model endpoints
• MLOps fundamentals: model versioning, pipeline automation, and CI/CD for ML with SageMaker Pipelines

🔬  Lab 1: ML Lifecycle on Amazon SageMaker

• Set up a SageMaker Studio environment and navigate the ML development workspace
• Load and explore a real dataset using SageMaker Data Wrangler
• Train a classification model using SageMaker built-in XGBoost algorithm
• Deploy a real-time inference endpoint and test predictions via the AWS console
• Evaluate model performance using SageMaker Experiments and compare model versions

2.1 Generative AI Concepts & Architecture

• What is Generative AI — difference from traditional ML, use cases, and business value
• Transformer architecture: attention mechanisms, encoders, decoders, and self-attention — conceptual understanding
• Large Language Models (LLMs): how they work, pre-training, fine-tuning, and the concept of model scale
• Tokenisation, embeddings, and vector representations — understanding how text becomes numbers for LLMs
• Foundation Models vs. Task-specific models — when to use each and trade-offs in cost and performance
• Multimodal AI: models that process text, images, audio, and video — AWS services and use cases

2.2 Prompt Engineering

• What is a prompt — components, structure, and the role of context in model output quality
• Zero-shot, one-shot, and few-shot prompting techniques with practical examples
• Chain-of-thought (CoT) prompting and tree-of-thought approaches for reasoning tasks
• Inference parameters: temperature, top-k, top-p (nucleus sampling), max tokens — effect on output
• Prompt injection risks and mitigation strategies in enterprise AI deployments
• System prompts and role-based prompting for business application development

2.3 AWS Generative AI Services

• Amazon Bedrock: overview, supported foundation models (Anthropic Claude, Meta Llama, Amazon Titan, Mistral, Cohere)
• Amazon Q: AWS’s enterprise AI assistant — use cases in code generation, business intelligence, and productivity
• AWS Trainium and AWS Inferentia: purpose-built AI chips for cost-efficient training and inference
• Amazon SageMaker JumpStart: pre-built ML solutions and foundation model deployment

🔬  Lab 2: Prompt Engineering with Amazon Bedrock

• Access Amazon Bedrock console and explore available foundation models
• Test zero-shot, one-shot, and few-shot prompts on a business summarisation task
• Experiment with temperature and top-p parameters and observe output variability
• Build a chain-of-thought prompt for a multi-step reasoning scenario
• Compare outputs across two different Bedrock foundation models for the same prompt

3.1 Selecting Foundation Models for Business Use Cases

• Model selection criteria: task type, context window, latency, cost, and language support
• Comparing models for text generation, summarisation, question answering, code generation, and translation
• Cost optimisation strategies: on-demand vs. provisioned throughput, batching, and caching
• Evaluating model outputs: human evaluation, automated metrics (BLEU, ROUGE, BERTScore), and benchmark datasets
• Business objective alignment metrics: task completion rate, user satisfaction, cost per interaction

3.2 Fine-Tuning and Model Customisation

• When to fine-tune vs. prompt engineer vs. use RAG — decision framework for enterprise use cases
• Fine-tuning approaches in Amazon Bedrock: instruction fine-tuning and continued pre-training
• Preparing training datasets for fine-tuning — data quality, format, and volume requirements
• Model evaluation after fine-tuning: comparing base model vs. fine-tuned model outputs

3.3 Retrieval-Augmented Generation (RAG)

• What is RAG and why it solves the hallucination problem — architecture overview
• RAG pipeline: Document Ingestion → Chunking → Embedding → Vector Store → Retrieval → Generation
• Amazon Bedrock Knowledge Bases: creating, populating, and querying a knowledge base
• Amazon OpenSearch Service as a vector store for enterprise RAG implementations
• Evaluating RAG pipeline quality: context relevance, faithfulness, and answer relevancy

3.4 Agentic AI and Amazon Bedrock Agents

• What are AI agents — autonomous task planning, tool use, and multi-step reasoning
• Amazon Bedrock Agents: creating agents, defining action groups, and connecting to APIs
• Amazon Bedrock Guardrails: content filtering, topic denial, PII redaction, and grounding checks
• Amazon Bedrock Data Automation: automated data extraction and document processing
• Amazon Bedrock Model Evaluation: automated and human-based evaluation workflows

🔬  Lab 3: Building a RAG Pipeline with Amazon Bedrock Knowledge Bases

• Upload enterprise documents (PDFs) to an Amazon S3 bucket as a knowledge source
• Create an Amazon Bedrock Knowledge Base and configure an OpenSearch Serverless vector store
• Configure chunking strategy and embedding model (Amazon Titan Embeddings)
• Query the knowledge base using Amazon Bedrock and compare RAG responses vs. base model responses
• Test Bedrock Guardrails: configure content filters and observe PII redaction in action

4.1 Responsible AI Principles

• AWS’s Responsible AI dimensions: Fairness, Explainability, Privacy, Robustness, Transparency, and Governance
• Understanding bias in AI/ML: types of bias (data bias, label bias, sampling bias) and their business impact
• Detecting and mitigating bias using Amazon SageMaker Clarify — bias metrics and pre/post-training analysis
• Explainability techniques: SHAP values, feature importance scores, and partial dependence plots in Clarify
• Human-in-the-loop (HITL) systems: when and how to incorporate human oversight in AI workflows
• Model cards and transparency documentation: communicating model limitations to stakeholders

4.2 Legal, Ethical & Regulatory Considerations

• AI regulations overview: EU AI Act, GDPR data privacy implications for AI, and US AI executive orders
• Intellectual property considerations in generative AI: training data copyright and output ownership
• Preventing harmful outputs: hallucination, toxicity, deepfakes, and disinformation risks
• Environmental impact of AI: energy consumption of large model training and sustainability considerations
• AWS’s shared responsibility model applied to AI workloads: what AWS manages vs. customer responsibility

🔬  Lab 4: Bias Detection and Explainability with Amazon SageMaker Clarify

• Configure a SageMaker Clarify bias detection job on a classification dataset
• Analyse pre-training bias metrics: Class Imbalance, Difference in Positive Proportions
• Run post-training bias analysis and interpret results in the SageMaker console
• Generate SHAP explainability report and identify top feature contributors to model predictions
• Review a sample Model Card and discuss transparency documentation best practices

5.1 Securing AI Workloads on AWS

• Identity and Access Management (IAM) for AI services: roles, policies, and least privilege for Bedrock and SageMaker
• Data encryption at rest and in transit: AWS KMS integration with SageMaker, Bedrock, and S3
• Network security for AI workloads: VPC endpoints, private model endpoints, and PrivateLink
• Securing training data: Amazon Macie for PII detection, S3 bucket policies, and access logging
• Audit logging for AI actions: AWS CloudTrail integration with Bedrock and SageMaker

5.2 Governance and Compliance for AI

• AWS AI governance framework: policies, model access controls, and cross-account Bedrock configurations
• Amazon Bedrock Model Access: enabling, restricting, and auditing foundation model usage
• AWS Config rules for compliance monitoring of AI resource configurations
• Data residency and sovereignty considerations for AI workloads in regulated industries
• Recognising governance and compliance regulations relevant to AI: HIPAA, SOC 2, PCI DSS implications
• AI incident management: monitoring, alerting, and responding to model performance degradation

🔬  Lab 5: AI Security & Governance on AWS

• Configure IAM roles and least-privilege policies for Amazon Bedrock model access
• Enable CloudTrail logging for Bedrock API calls and query logs in CloudWatch
• Set up Amazon Macie on an S3 bucket containing training data and review PII findings
• Configure a VPC endpoint for Amazon SageMaker to restrict network access
• Review AWS Config rules applicable to SageMaker endpoints and simulate a compliance check

Capstone Project: End-to-End AI Solution on AWS

Design and build a complete AI solution for a simulated business scenario — selecting the appropriate AWS AI services, configuring an Amazon Bedrock-powered RAG pipeline with Guardrails, applying responsible AI controls using SageMaker Clarify, and documenting security and governance considerations. Present your architecture diagram, service selection rationale, and model evaluation results to the instructor and batch cohort.

Certification Preparation — AIF-C01 Focus Areas

• Domain 1: AI/ML Fundamentals — learning types, ML lifecycle, and AWS ML services
• Domain 2: Generative AI — LLMs, foundation models, prompt engineering, Bedrock models
• Domain 3: Foundation Model Applications — RAG, Agents, Guardrails, model selection and evaluation
• Domain 4: Responsible AI — bias, explainability, fairness, Clarify, ethical and regulatory frameworks
• Domain 5: Security & Governance — IAM, KMS, CloudTrail, Macie, compliance controls
• Full-length mock tests: 65 questions, 120 minutes, scenario-based questions aligned to AIF-C01 pattern
• Case study walkthroughs: retail AI recommendation, healthcare NLP, financial fraud detection scenarios