AWS Machine Learning Engineer Associate

Collect and store data (S3, EFS, RDS, DynamoDB, OpenSearch; storage decisions) · Streaming ingestion (Kinesis, Data Firehose, Apache Flink, Kafka) · Data formats by access pattern (Parquet, ORC, JSON, CSV) · Merging sources (AWS Glue, Apache Spark) · Scalable vector databases (OpenSearch, RDS pgvector, S3 Vectors) · SageMaker Feature Store ingestion and management · Transformation and feature engineering (Glue, DataBrew, EMR, Data Wrangler) · Embedding models and text pre-processing · RAG document preparation (chunking, metadata extraction) · Masking, redaction, and anonymization · Data for FM fine-tuning, continued pre-training, and distillation · Data quality validation, labeling, bias metrics, and class imbalance

Selecting Bedrock FMs and fine-tuning strategies · Choosing ML/GenAI models, algorithms, and solution templates · Custom vs managed vs pre-trained vs FM trade-offs · RAG architecture patterns · Performance, latency, training-time, and cost trade-offs · AWS AI services (Textract, Rekognition, Comprehend, Transcribe) · SageMaker AI built-in algorithms, script mode, and automatic model tuning · Training-time reduction, overfitting, catastrophic forgetting, ensembles · Prompt engineering and retrieval/embedding optimization · Reproducible experiments (MLflow, Bedrock evaluations, Prompt Management) · Evaluation metrics, explainability, convergence debugging, shadow variants · Human evaluation, NLP metrics (BLEU, ROUGE, BERTScore), LLM-as-a-judge, RAG evaluation

Compute environments, deployment targets, and orchestrators · Real-time, serverless, asynchronous, and batch inference · FM deployment options and Bedrock Custom Model Import · Agents, agent tools, and agent communication protocols (MCP, A2A) · RAG configuration (retrieval strategies, reranking) · On-demand vs provisioned resources; infrastructure as code (CloudFormation, CDK) · Containers, VPC-attached endpoints, programmatic deployment, auto scaling · Bedrock knowledge bases and retrieval pipelines · Agent state management, GPU scaling, agentic workflow infrastructure · Deployment strategies and rollback; CodePipeline, CodeBuild, CodeDeploy · Training/inference job orchestration, automated testing, retraining · Model, prompt, and agent version management; RAG refresh orchestration

Production model monitoring and drift detection · A/B testing and agent performance monitoring · FM performance monitoring (Bedrock evaluations, CloudWatch generative AI observability) · Instance selection, troubleshooting tools, dashboards, and capacity · Cost management tools and purchasing options · FM inference cost and agent resource consumption · Securing CI/CD (code and image scanning) · Least-privilege IAM for ML and AI artifacts · Monitoring, auditing, and compliance logging (CloudTrail, Config) · Network isolation (VPCs, subnets, security groups) · AI security risks and FM credential types · Bedrock Guardrails and sensitive data protection