AWS Machine Learning Specialty (AWS) study plan

Set up a personal AWS sandbox account with budget alerts and an S3 bucket for lab artifacts

Load a public tabular dataset into Amazon S3 and query it in place with Athena to practice schema-on-read exploration

Compute summary statistics and detect nulls, duplicates, and outliers on that dataset using Athena SQL and a pandas notebook

Build a SageMaker Studio domain with a single-user profile and confirm notebook and Data Wrangler access

Compare distributions across feature groups using histograms, box plots, and correlation matrices in SageMaker Data Wrangler

Apply imputation, encoding, scaling, and train/test splitting transforms and export the resulting flow to a processing job

Ingest streaming records into Kinesis Data Streams and land them in S3 in partitioned Parquet via Firehose

Run an AWS Glue crawler over the partitioned data and validate the Data Catalog schema

Train at least three built-in SageMaker algorithms (for example XGBoost, Linear Learner, and K-Means) on the prepared dataset

Tune hyperparameters with a SageMaker automatic model tuning job and interpret the objective metric curve

Configure a SageMaker Processing job to run a reproducible feature-engineering script outside the notebook

Explain when to choose built-in algorithms versus bring-your-own-container versus framework containers

Diagnose underfitting and overfitting by comparing training, validation, and test metrics for each model

Evaluate classification models with precision, recall, F1, ROC-AUC, and confusion matrices at multiple thresholds

Evaluate regression models with RMSE, MAE, and residual plots and identify heteroscedasticity

Detect and quantify class imbalance, then apply resampling or class-weighting and re-measure

Identify and mitigate leakage by auditing the feature set against the split strategy

Deploy the chosen model to a real-time SageMaker endpoint and invoke it from a Python client

Configure a batch transform job for the same model and compare cost and latency against real-time hosting

Register the model in SageMaker Model Registry with a model card and approval status

Set up a multi-model endpoint or serverless endpoint and explain the tradeoffs of each hosting option

Build an automated SageMaker Pipelines workflow covering processing, training, evaluation, and conditional registration

Add a quality-gate condition step that blocks registration when the evaluation metric falls below a threshold

Configure CloudWatch alarms and a Model Monitor baseline for data drift and model quality drift

Automate endpoint updates using a deployment guardrail such as shadow or linear traffic shifting

Harden the solution by scoping IAM roles to least privilege and moving secrets into Secrets Manager

Encrypt S3, EBS, and endpoint traffic with KMS keys and verify the settings in each service console

Instrument the pipeline with CloudWatch Logs and metrics and build a dashboard for job failures and latency

Estimate cost for training, hosting, and storage and identify at least three cost-reduction levers

Take two full-length timed practice exams and log every missed question by domain

Re-derive the reasoning for each missed item against official service documentation rather than memorizing answers

Review weak areas in EDA and modeling, the two heaviest domains, with targeted hands-on re-labs

Rehearse explaining algorithm selection, evaluation tradeoffs, and deployment patterns aloud in under two minutes each