AWS AI & SageMaker
Amazon Web Services is the world's largest cloud provider, with the most mature and comprehensive set of AI/ML services. SageMaker is its flagship ML platform — handling everything from data preparation to model training, deployment, and monitoring.
AWS AI Service Landscape
End-to-end ML platform: notebooks, training, deployment, monitoring
Managed LLM API: Claude, Llama, Titan, Stable Diffusion. No GPU management.
Pre-built computer vision: object detection, face recognition, content moderation
NLP service: sentiment, entities, key phrases, language detection
Speech-to-text and text-to-speech services
Data preparation and SQL analytics on S3 data lakes
SageMaker: The ML Lifecycle
Web-based IDE. Jupyter notebooks with managed compute. No local setup.
Visual data preparation: clean, transform, and visualise S3 data.
Managed container-based training. Auto-provisions GPUs, handles checkpointing.
Track hyperparameters, metrics, and artefacts across training runs.
Version models, track lineage, approve/reject for deployment.
Deploy models as real-time or batch REST APIs. Auto-scaling built-in.
Detect data drift and model quality degradation in production.
Launch a Training Job
import sagemaker
from sagemaker.pytorch import PyTorch
role = sagemaker.get_execution_role()
sess = sagemaker.Session()
# Upload training data to S3
s3_data = sess.upload_data('data/', bucket='my-bucket', key_prefix='training')
# Define estimator
estimator = PyTorch(
entry_point='train.py', # Your training script
role=role,
framework_version='2.1',
py_version='py310',
instance_type='ml.p3.2xlarge', # V100 GPU, $3.82/hr
instance_count=1,
hyperparameters={
'epochs': 50,
'learning-rate': 0.001,
}
)
# Start training (provisions GPU, runs, saves model to S3)
estimator.fit({'training': s3_data})Deploy a Model Endpoint
# Deploy trained model as REST API
predictor = estimator.deploy(
initial_instance_count=1,
instance_type='ml.c5.xlarge', # CPU for inference, cheaper
endpoint_name='my-model-prod'
)
# Make predictions
import json
result = predictor.predict({'inputs': [[5.1, 3.5, 1.4, 0.2]]})
print(result)
# Clean up (endpoints cost money even when idle!)
predictor.delete_endpoint()AWS Bedrock — Managed LLM API
Bedrock gives you API access to foundation models (Claude, Llama, Titan) without managing any infrastructure. Pay per token, scale to zero automatically.
import boto3
import json
bedrock = boto3.client('bedrock-runtime', region_name='us-east-1')
response = bedrock.invoke_model(
modelId='anthropic.claude-3-5-sonnet-20241022-v2:0',
body=json.dumps({
"anthropic_version": "bedrock-2023-05-31",
"max_tokens": 512,
"messages": [{"role": "user", "content": "Summarise this document..."}]
})
)
result = json.loads(response['body'].read())
print(result['content'][0]['text'])Cost Optimisation Tips
Save up to 90% on training jobs. SageMaker handles interruption and checkpointing automatically.
ml.p3.2xlarge (~$3.82/hr) for training; ml.c5.xlarge (~$0.19/hr) for CPU inference.
Endpoints bill 24/7. Use Lambda to auto-scale to zero or delete during off-hours.
For intermittent traffic — only charges per prediction. Zero cost when idle.
Frequently Asked Questions
What IAM permissions does SageMaker need?
SageMaker needs an execution role with: AmazonSageMakerFullAccess (for SageMaker APIs), S3 read/write access for your data bucket, ECR access if using custom containers, and CloudWatch Logs access. Never use AdministratorAccess in production — follow least privilege.
How do I version my models in SageMaker?
Use the SageMaker Model Registry. Register models after training with estimator.register(). Models go through an approval workflow (Pending → Approved/Rejected) before being deployed to production endpoints. Model lineage tracks which training job and data version produced each model.