Amazon SageMaker

End-to-end managed ML: collect and prepare data, build and train, deploy and monitor, all in one place. It’s for developers and data scientists building custom models (vs the pre-trained AI services).

Built-in algorithms

Supervised (linear regression and classification, KNN), unsupervised (PCA for dimensionality reduction, K-means, anomaly detection), text (NLP, summarization), image (classification, detection). Plus DeepAR for time-series forecasting (RNN-based).

Deployment and inference

• Real-time: ms-to-seconds latency, payloads up to 25 MB. Instant predictions for web and mobile.
• Serverless: same latency but cold starts are OK, up to 4 MB. Sporadic traffic, no infra.
• Asynchronous: near real time, up to 1 GB and a 1 hour max, S3 in and out. Big payloads, long processing.
• Batch transform: minutes to hours, 100 MB per mini-batch, S3 in and out. Whole datasets at once.

Automatic Model Tuning (AMT): give it an objective metric and it picks hyperparameter ranges, the search strategy, and early stopping. Saves money on bad configs.

The tool zoo (know the one-liners)

• Studio: the unified interface for everything below.
• Data Wrangler: prepare and transform data, feature engineering, visual + SQL, data quality.
• Feature Store: a central store for ML features, discoverable, reusable.
• Clarify: evaluate and compare FMs, explain predictions, detect bias in data and models.
• Ground Truth: humans label data and grade models, RLHF. Ground Truth Plus is the managed labeling version.
• Model Cards: document a model: intended use, risk rating, training details.
• Model Dashboard: all models in one portal, flagging threshold violations.
• Model Monitor: production quality monitoring with drift alerts.
• Model Registry: version, catalog, approve, and share models.
• Pipelines: CI/CD for ML: automate build, train, test, and deploy across steps (Processing, Training, Tuning, AutoML, ClarifyCheck, QualityCheck).
• Role Manager: persona-based access (data scientist, MLOps engineer).
• JumpStart: the model hub: pre-trained FMs (Hugging Face, Meta, Stability…) and pre-built solution templates, deployed into SageMaker under your control.
• Canvas: no-code visual ML with AutoML, plus ready-made models from Rekognition, Comprehend, and Textract.
• MLFlow on SageMaker: managed MLFlow tracking servers for experiments.

Clarify bias notes: it detects skew (data over-representing middle-aged people, say); fix imbalanced classes with Data Wrangler augmentation. The bias types worth naming: sampling, measurement, observer, confirmation.

Network isolation mode runs training containers with no outbound internet, not even S3.