AIF-C01: Cheatsheet
Exam Day Reference
Print this or review 5 minutes before the exam.
AWS AI Services Quick Lookup
| Need | Service |
|---|---|
| Foundation models | Amazon Bedrock |
| Business assistant | Amazon Q |
| Code suggestions | Amazon CodeWhisperer |
| Full ML platform | Amazon SageMaker |
| Image/video analysis | Amazon Rekognition |
| Extract text | Amazon Textract |
| Sentiment analysis | Amazon Comprehend |
| Translation | Amazon Translate |
| Speech-to-text | Amazon Transcribe |
| Text-to-speech | Amazon Polly |
| Chatbot | Amazon Lex |
| Recommendations | Amazon Personalize |
| Fraud detection | Amazon Fraud Detector |
| Search | Amazon Kendra |
| Human review | Amazon A2I |
Bedrock Models
| Model | Provider | Best For |
|---|---|---|
| Claude | Anthropic | Long docs (200K), reasoning, code |
| Titan Text | Amazon | Cost-effective, simple tasks |
| Titan Embeddings | Amazon | RAG, semantic search |
| Titan Image | Amazon | Image generation |
| Jurassic | AI21 Labs | Multilingual text |
| Command | Cohere | Text generation |
| Llama 2 | Meta | Open-source, customizable |
| Stable Diffusion | Stability AI | Image generation |
ML Metrics (PARC)
Precision = TP / (TP + FP)
- Use when: False positives costly (spam filter)
Accuracy = (TP + TN) / Total
- Use when: Balanced classes
Recall = TP / (TP + FN)
- Use when: False negatives costly (medical diagnosis)
Confusion Matrix
- TP, TN, FP, FN
Model Fit Issues
| Issue | Performance | Solution |
|---|---|---|
| Underfitting | Poor on train AND test | More complex model |
| Good Fit | Good on both | ✅ Goal |
| Overfitting | Great on train, poor on test | More data, regularization |
RAG Workflow (5 Steps)
- User asks question
- Convert question to embedding (vector)
- Search vector database for similar content
- Retrieve relevant documents
- LLM generates answer with context
AWS Vector DB: Amazon OpenSearch Service
Prompt Engineering Ladder
Zero-Shot: No examples
- "Translate to French: Hello"
Few-Shot: 3-5 examples
- Show input/output pairs
Chain-of-Thought: Show reasoning
- "Show your work"
Customization Approaches
| Approach | When | Cost | Effort |
|---|---|---|---|
| Prompt Engineering | First choice | Low | Low |
| RAG | Current/private data | Medium | Medium |
| Fine-Tuning | Domain-specific | High | High |
Order: Always try Prompt → RAG → Fine-Tuning
Responsible AI (FEPST)
- Fairness — Avoid bias, treat all equitably
- Explainability — Understand decisions
- Privacy — Protect data (encryption, IAM)
- Safety — Prevent harmful outputs
- Transparency — Document capabilities/limits
Responsible AI Tools
| Need | AWS Tool |
|---|---|
| Detect bias | SageMaker Clarify |
| Monitor drift | SageMaker Model Monitor |
| Human review | Amazon A2I |
Security Checklist
- ✅ Encrypt at rest (AWS KMS)
- ✅ Encrypt in transit (TLS/SSL)
- ✅ IAM policies for access
- ✅ Input validation
- ✅ Monitor for prompt injection
- ✅ Bedrock data stays private (not used for training)
Compliance
| Regulation | What It Means |
|---|---|
| HIPAA | Healthcare data (use HIPAA-eligible services) |
| GDPR | EU data privacy (right to explanation) |
| SOC 2 | Security controls |
| PCI DSS | Payment card data |
Learning Types
| Type | Data | Example |
|---|---|---|
| Supervised | Labeled (X, y) | Spam detection |
| Unsupervised | Unlabeled | Customer segmentation |
| Reinforcement | Rewards | Game AI |
The ML Pipeline
- Business Goal ➡️ 2. Data Prep ➡️ 3. Model Training ➡️ 4. Evaluation ➡️ 5. Deployment (Inference) ➡️ 6. Monitoring
Inference Types
- Real-time: Low latency, immediate (e.g., fraud check at checkout).
- Batch: High volume, delayed (e.g., monthly product recommendations).
GenAI Limitations
- ❌ Hallucinations (generates false info)
- ❌ Training cutoff (no current events)
- ❌ Context limits (can't process very long docs)
- ❌ Bias (reflects training data)
- ❌ No real-time data (use RAG)
Mitigation: RAG, guardrails, human review
Last Updated: 2026-02-05