Amazon AIF-C01: ML Data Engineering — Study Guide
Part of the AWS AI Practitioner AIF-C01 — Study Guide. Practice with verified answers in the Amazon exam hub, or take timed practice tests on ExamRoll.io.
Data collection, labeling, versioning, and governance
Collecting representative, auditable datasets is the foundational step for all ML. Use Amazon S3 as the central durable store with S3 Versioning and object lock enabled to preserve raw inputs and capture provenance. For structured ingestion and cataloging, register datasets in the AWS Glue Data Catalog and apply fine-grained access control using Lake Formation. When human labeling is required, Amazon SageMaker Ground Truth provides built-in workflows, annotation UI, and active learning loops that reduce labeling cost while creating metadata for labeler agreement and confidence. Common traps include collecting biased or nonrepresentative samples, failing to record lineage and labeling rules, and insufficient label quality checks; mitigate these by storing labeling audits, blind-relabel subsets, and using label consolidation heuristics. Privacy and compliance decisions drive architecture: use S3 server-side encryption with KMS-managed CMKs, restrict access with IAM policies and VPC endpoints (AWS PrivateLink), and perform PII discovery and redaction with Amazon Comprehend before sharing data for model training. For long-running programs, capture dataset snapshots, tag with dataset IDs, and track experiments with Amazon SageMaker Experiments or an external tool like DVC to enable reproducible training and regulatory reporting.
Preprocessing, feature engineering, and EDA
Transformations and features determine model capacity to generalize. Use AWS Glue or Amazon SageMaker Data Wrangler to profile, clean, and normalize data, and perform iterative exploratory data analysis with Amazon QuickSight or Jupyter notebooks hosted in Amazon SageMaker Studio. For production feature management, adopt Amazon SageMaker Feature Store to ensure consistent offline and online feature computation and to avoid train/serve skew; store raw and derived features separately and record feature creation logic. Time-series and streaming pipelines should leverage Amazon Kinesis or AWS Glue streaming ETL for low-latency feature refresh. Typical traps include data leakage—where future information leaks into training—improper handling of missing values, and mismatch between offline training features and online serving features. Decision criteria when designing pipelines hinge on freshness versus cost: choose batch ETL for heavy historical recomputations, streaming for near-real-time personalization, and hybrid architectures when low-latency online features are required. Automate validation checks and schema enforcement in Glue or as part of SageMaker Processing jobs to catch schema drift early.
Embeddings, augmentation, synthetic data, and datasets for foundation models
Embeddings convert text, images, or multimodal inputs into dense vectors that capture semantic relationships; they power semantic search, retrieval-augmented generation, and clustering. Build a retrieval-augmented generation (RAG) pattern where you generate embeddings with Amazon Bedrock or SageMaker-hosted encoders, store vectors in Amazon OpenSearch Service (k-NN), Amazon Kendra, or a managed vector store, and retrieve context to condition the foundation model. Data augmentation and synthetic data generation are practical when real examples are scarce: use generative models in SageMaker to create paraphrases, image transforms (via Albumentations or SageMaker Processing), or privacy-preserving synthetic records. Beware of over-reliance on synthetic data—poor fidelity can introduce distribution shift and overfit models to artifacts. For FM training datasets, curate high-quality exemplars, canonicalize formats with prompt templates, and version every dataset snapshot in S3. For private data used with third-party FMs, prefer a private fine-tuning path (bring compute into a VPC) or deploy retrieval-only pipelines that only send non-sensitive context to the FM while keeping source documents in a secure vector store; apply redaction and token-level masking to prevent leakage. Prompt engineering and instruction templates (system prompts) are critical to shape model responses; tune temperature, top_k, or top_p according to the required determinism and creativity.
Evaluation, deployment, monitoring, and lifecycle management
Evaluation must be explicit, using metrics aligned to business objectives: regression tasks use RMSE or MAE, binary classification evaluates precision/recall/F1 and ROC AUC, and summarization uses ROUGE variants. Conduct validation and cross-validation during the evaluation phase and hold out a blind test set for final acceptance. Deploy using Amazon SageMaker endpoints (real-time or serverless inference) or Amazon Bedrock for managed FM serving; optimize for latency by selecting smaller distilled models, enabling multi-model endpoints, or using SageMaker Serverless Inference for unpredictable traffic. Monitor performance and data drift in production with Amazon SageMaker Model Monitor and set up alerts for metric degradation; employ canary or blue/green deployments to limit risk. Access control for model usage is enforced using IAM role policies, resource-level permissions, and network isolation. Practitioner traps include choosing the wrong metric (accuracy for imbalanced classes), ignoring concept drift, and failing to instrument training data and inference logs for auditability. Use CI/CD for ML with SageMaker Pipelines to standardize retraining cadence, keep dataset and model versioning consistent, and maintain a defensible audit trail for compliance.
Practical Problem: Use-Case Scenario
Scenario: BrightCart, an online grocery chain, is modernizing its on-prem inventory system into AWS and wants a generative AI chatbot that answers customer questions and returns live inventory locations while protecting customer and inventory data in accordance with compliance rules. Their AWS AI environment includes S3 for datasets, Amazon RDS for transactional inventory, SageMaker Studio for development, Bedrock access for foundation models, and VPC networking.
Challenge: Build a private, low-latency RAG-enabled chatbot that retrieves current inventory and avoids exposing sensitive data or triggering hallucinations.
Recommended Approach:
- Provision a private VPC and configure AWS PrivateLink endpoints for Bedrock and SageMaker; store canonical product documents and inventory snapshots in S3 with server-side encryption (KMS) and enable S3 versioning.
- Continuously compute embeddings from product docs with a Bedrock encoder or SageMaker model in a private subnet, and store vectors in Amazon OpenSearch Service with k-NN for fast nearest-neighbor retrieval; keep live inventory in Amazon RDS and provide a secure runtime connector for retrieval.
- Implement a retrieval-augmented pipeline where the application first queries OpenSearch for context, then calls Bedrock with a system prompt template that includes retrieved context and explicit safety instructions; sanitize user inputs via input validation and use Amazon Comprehend to detect PII and redact before retrieval.
- Deploy the chatbot behind an Application Load Balancer, serve the FM via Bedrock with an API gateway in the VPC, enable logging to CloudWatch with restricted access, and monitor model responses and drift with SageMaker Model Monitor and periodic human audits.
Rationale: This approach uses a RAG pattern to minimize sensitive data exposure, leverages private networking and KMS for compliance, separates live inventory from model context for accuracy, and applies monitoring and human-in-the-loop checks to control hallucinations and ensure ongoing reliability.
← AI Security · All domains · Model Training →
Practice these questions → · Timed practice on ExamRoll.io →
Pass the whole exam — not just this question
You found this answer. Get every verified question and explanation in one place, and save hours of prep. Free to start.
Pass your exam →