Exam Guide & Traps

← Domain 5 · Cheatsheet →

---

Exam Traps

Trap 1: Bedrock Model Confusion

Know which model for which scenario:

If question mentions...	Think...
Long documents (200K tokens)	Anthropic Claude
Cost-effective text generation	Amazon Titan Text
Embeddings for RAG	Amazon Titan Embeddings
Image generation	Stable Diffusion XL
Multilingual content	Cohere or AI21 Jurassic

Trap 2: RAG vs Fine-Tuning vs Prompt Engineering

Common mistake: Choosing fine-tuning when RAG or prompting would work.

Correct order:

1. Prompt Engineering (try first - cheapest, fastest)
2. RAG (need current/private data)
3. Fine-Tuning (only if above methods insufficient)

Examples:

• ❌ "Fine-tune the model to access current news"

• ✅ "Use RAG to provide current news as context"

• ❌ "Fine-tune for better output format"

• ✅ "Use few-shot prompting with examples"

Trap 3: ML Metrics Selection

Scenario	Metric	Why
Spam filter	Precision	False positives costly (important emails marked spam)
Medical diagnosis	Recall	False negatives costly (miss diseases)
Balanced dataset	Accuracy	Classes are equal
Imbalanced data	Precision/Recall	Accuracy misleading

Trap 4: Responsible AI Tools

Need	AWS Service
Detect bias in data/models	SageMaker Clarify
Monitor model drift	SageMaker Model Monitor
Human review workflows	Amazon A2I
Explain predictions	SageMaker Clarify

Trap 5: Security Best Practices

When asked about securing AI systems:

• ✅ Encrypt at rest (AWS KMS, S3 encryption)
• ✅ Encrypt in transit (TLS/SSL)
• ✅ IAM policies for access control
• ✅ Input validation and sanitization
• ✅ Monitor for prompt injection attacks

Common trap: Forgetting that Bedrock data does not train public models.

---

Decision Quick Reference

Bedrock vs SageMaker Quick Decision Tree

Need to use AI/ML?
├─ Want pre-built models via API?
│  ├─ GenAI/LLMs? → Amazon Bedrock ⭐
│  └─ Specific tasks? → AI Services (Rekognition, Comprehend, etc.)
│
└─ Want to build/train custom models?
   └─ Full ML lifecycle control? → SageMaker ⭐

Decision Quick Reference

"Which AWS AI service?"

Extract text from documents → Amazon Textract
Analyze sentiment → Amazon Comprehend
Translate languages → Amazon Translate
Speech-to-text → Amazon Transcribe
Text-to-speech → Amazon Polly
Build chatbot → Amazon Lex
Recommendations → Amazon Personalize
Fraud detection → Amazon Fraud Detector
Search documents → Amazon Kendra
Access foundation models → Amazon Bedrock
Full ML platform → Amazon SageMaker

"RAG, Fine-Tuning, or Prompt Engineering?"

Simple task, common format → Prompt Engineering (zero-shot)
Custom format, few examples → Prompt Engineering (few-shot)
Need current data → RAG
Need private/proprietary data → RAG
Have 100+ training examples → Fine-Tuning
Domain-specific terminology → Fine-Tuning
Budget constrained → Prompt Engineering

"Which type of learning?"

Have labeled data (X and y) → Supervised Learning
No labels, find patterns → Unsupervised Learning
Learn through rewards → Reinforcement Learning

"How to address concerns about..."

Hallucinations → Use RAG, implement guardrails, human review
Data privacy → Bedrock doesn't train on your data, encryption
Bias → Use SageMaker Clarify, diverse training data
Cost → Choose appropriate model size, optimize prompts, cache
Model drift → SageMaker Model Monitor, regular retraining

---

Exam Day Reminders

Think Like This

For Bedrock questions:

• Longest context? → Claude (200K)
• Cheapest? → Titan
• Embeddings? → Titan Embeddings
• Images? → Stable Diffusion

For ML questions:

• False positives bad? → Precision
• False negatives bad? → Recall
• Model too simple? → Underfitting
• Memorized training data? → Overfitting

For Responsible AI:

• Detect bias? → SageMaker Clarify
• Monitor drift? → SageMaker Model Monitor
• Human review? → Amazon A2I

For RAG:

• Always: Embeddings → Vector DB → Retrieve → Generate
• AWS vector DB: Amazon OpenSearch Service

---

← Domain 5 · Cheatsheet →