AWS AI and ML Services

Overview

AWS provides a layered AI/ML stack — from fully managed AI services (no ML knowledge needed) through SageMaker (full ML lifecycle control) to Bedrock (managed foundation model access). This page synthesizes the key services and decision points for the AWS AI Practitioner (AIP-C01) certification.

---

Amazon SageMaker

Full ML lifecycle platform for building, training, and deploying custom models.

Training

• Hyperparameter tuning — automatic optimization under SageMaker model training
• Warm start — reuse previous tuning jobs as starting points
• EnableNetworkIsolation=true blocks internet access during training
• Input modes:
  • File mode — copies entire dataset from S3 before training (slower for large data)
  • Fast File mode — streams data directly from S3 on demand
  • FSx for Lustre — high-performance reads for large datasets

Key Components

Component	Purpose
Canvas	No-code ML model building with visualization (#nocode)
JumpStart	Pretrained open-source models for quick starts (#nocode)
Data Wrangler	Data preparation and transformation (now Canvas)
DeepAR	Time-series forecasting using RNNs
Model Registry	Centralized model catalog
Model Cards	Document model characteristics
Clarify	Detect bias in data
Model Monitor	Continuous monitoring for drift; baseline from training data

Inference/Deployment Options

Type	Latency	Max Size	Use Case
Real-time endpoints	Sub-millisecond	25 MB	Continuous low-latency inference
Asynchronous endpoints	Up to 15 min	1 GB	Long-running workloads, GPU-supported
Serverless inference	Auto-scaling	—	Variable traffic, no GPU
Batch transform	Offline	Large batches	Offline processing, results stored in S3

• Multi-model endpoints deploy multiple models on a single endpoint
• Can combine with Lambda, Bedrock provisioned throughput, or hybrid solutions

---

Amazon Bedrock

Fully managed, serverless access to third-party foundation models for generative AI.

Key Features

• On-demand API access to foundation models
• Import models trained/fine-tuned in SageMaker
• Cross-Region Inference for model availability
• Bedrock Data Automation (BDA) — process documents, images, video, audio
• Bedrock AgentCore — build, deploy, and operate agents at scale
• Knowledge Bases — managed RAG implementation
• Streaming support via BedrockRuntimeClient for real-time feedback

Bedrock KB RAG vs Kendra-based RAG

• Bedrock Knowledge Bases: vector-based retrieval integrated with foundation models
• Kendra: managed semantic search using NLP (non-vector, contextual ranking)

---

Other AI Services

Service	Purpose
Comprehend	NLP — key phrases, sentiment, entities, PII detection
Comprehend Medical	Detect PHI in medical text
Macie	Discover/protect PII in S3 using ML
Transcribe	Speech-to-text with auto data masking
Translate	Language translation
Textract	Image/document to text (OCR)
Lex	Chatbot building
Polly	Text-to-speech
Kendra	Semantic search with NLP and GenAI Index
QuickSight	BI dashboards with ML insights
Pinpoint	Marketing communications (email, SMS, push)

Agent Frameworks

• Strands Agents — SDK for autonomous AI agents
• Agent Squad — multi-agent orchestration

---

ML Concepts and Techniques

Fine-Tuning

• LoRA (Low-Rank Adaptation) — parameter-efficient fine-tuning with small trainable matrices
• Full fine-tuning — adjusts all parameters, better performance but more resources

Model Optimization

• Quantization — reduce numerical precision of weights/activations
• Pruning — remove unnecessary weights or neurons
• Knowledge distillation — transfer from large teacher to small student model

Search and Retrieval

• k-NN — exact nearest neighbor, computationally intensive, best for small datasets
• ANN (Approximate Nearest Neighbor) — HNSW algorithm, fast with near-optimal accuracy for large vector DBs

Evaluation

• LLM-as-a-judge — use one model to evaluate another’s outputs (bias, toxicity, guideline adherence)

Patterns

• ReAct — alternate between reasoning and taking actions
• Chain of Thought (CoT) — step-by-step reasoning
• Circuit Breaker — temporarily block access to faulty service after failures
• Exponential backoff with jitter — intelligent retry spacing adapting to system stress
• Recursive summarization — progressively condense long content

Data Techniques

• Data augmentation: rotation, flipping, scaling
• SMOTE: synthetic minority oversampling for imbalanced datasets
• L2 regularization (ridge regression): prevent overfitting
• Embedding drift: monitor when embedding model changes

---

Infrastructure Decision Points

Compute Location

Option	Use Case
Region	Standard workloads
Local Zones	Low-latency near population centers
Wavelength	5G mobile edge computing
Outposts	On-premises AWS infrastructure
Snowball	Large data migration

Networking

• PrivateLink — one-way VPC-to-VPC connection
• Transit Gateway — hub connecting multiple VPCs and on-premises
• Direct Connect — dedicated on-premises to AWS connection
• Site-to-Site VPN — encrypted tunnel (e.g., for Glue)

Data and Integration

• Lake Formation — centralized data governance for analytics/ML
• Glue — ETL, PII detection in pipelines, Data Catalog
• AppFlow — SaaS-to-AWS data synchronization
• EventBridge — complex event routing (vs SQS for simple queuing)
• Step Functions with Dynamic Choice — visual routing for document processing

Security

• SSE-KMS vs SSE-S3 — KMS provides more logs and fine-grained control
• Verified Permissions — fine-grained authorization service
• CloudWatch Anomaly Detection — detect token burst patterns in Bedrock apps
• API Gateway Request Validators — validate size and fields

---

Related

• Cloud & AWS Infrastructure
• AI & Machine Learning
• Serverless
• Retrieval-Augmented Generation (RAG)

---

Source References

• AIP C01 Prepare Notes