Question 1

What is Amazon SageMaker?

Accepted Answer

SageMaker is AWS's managed machine-learning platform covering the whole lifecycle: Studio for development, Data Wrangler and Processing for data preparation, Feature Store, training jobs with automatic model tuning, a model registry, endpoints for real-time and batch inference, Pipelines for orchestration, and Model Monitor for drift. The value to name is that it removes infrastructure management from each lifecycle stage — training clusters are provisioned and torn down per job rather than run continuously. That per-job model is what makes it cheaper than a standing GPU cluster for intermittent training.

Question 2

What is the difference between real-time, serverless, asynchronous and batch inference?

Accepted Answer

A real-time endpoint serves low-latency requests from always-running instances you pay for continuously. Serverless inference scales to zero and bills per invocation, suiting intermittent traffic with tolerance for cold starts. Asynchronous inference queues requests for large payloads or long processing. Batch transform runs over a dataset and terminates. The cost consequence is the practical point: a model used only for nightly scoring should use batch transform, not a persistent endpoint. Naming that a persistent endpoint bills continuously regardless of traffic is what makes the selection concrete.

Question 3

What is SageMaker Pipelines?

Accepted Answer

Pipelines defines an ML workflow as a directed graph of steps — processing, training, evaluation, conditional registration, deployment — with parameters, caching of unchanged steps, lineage tracking and integration with the model registry, executed as a managed service. Step caching changes the economics of iteration: re-running after changing only the training step skips expensive data preparation. Lineage — knowing exactly which data and code produced a deployed model — is the governance argument that matters in regulated environments.

Question 4

What is a feature store and why does it exist?

Accepted Answer

A feature store centralises the computation, storage and serving of features so the same definition is used for training and for online inference, with an online store for low-latency lookup and an offline store for historical training data. It prevents features being computed one way in a training query and another way in production code. That mismatch is training-serving skew, and naming it explicitly is what makes the answer land. The offline store also gives point-in-time correct joins, which prevents target leakage from using feature values that would not have been available at prediction time.

Question 5

What is training-serving skew and how do you prevent it?

Accepted Answer

A mismatch between the data distribution or feature computation at training time and at serving time, causing the deployed model to perform worse than evaluation suggested. Prevent it with a feature store or shared transformation code used by both paths, schema validation on serving inputs, and monitoring the live feature distribution against the training baseline. The subtle variant is timing: computing a feature from data unavailable at prediction time — target leakage — makes the model look excellent in evaluation and useless in production. Naming leakage as the related failure shows depth.

Question 6

What is SageMaker Model Monitor?

Accepted Answer

Model Monitor captures inference requests and responses and compares them against a baseline computed from training data, detecting data quality issues, feature drift, model quality degradation where labels arrive, and bias drift, raising CloudWatch alarms when thresholds are exceeded. The distinction between skew — serving data differs from training data — and drift — serving data changes over time — is what interviewers check. Both degrade a model silently, because accuracy is unobservable in production until labels arrive weeks later, which is why distribution monitoring is the practical proxy.

Question 7

What is MLOps and what does a mature AWS setup look like?

Accepted Answer

MLOps applies engineering discipline to machine learning: version-controlled data, code and models; automated reproducible training pipelines; automated evaluation gates before promotion; a model registry with lineage; controlled deployment with traffic shifting; and continuous monitoring with retraining triggers. On AWS that is SageMaker Pipelines, Model Registry, Feature Store and Model Monitor, wired together with CodePipeline and IaC. The maturity ladder — manual notebooks, then automated training, then automated retraining triggered by monitoring — is a good framing. The key detail is that a retrained model must pass the same evaluation gate as a manual one, or automation just deploys worse models faster.

Question 8

How do you deploy a model update safely?

Accepted Answer

Register the new version, then use a SageMaker deployment guardrail — blue/green with canary or linear traffic shifting — monitored against CloudWatch alarms with automatic rollback. Shadow testing sends a copy of live traffic to the new variant without serving its responses, so you compare behaviour before any user is affected. Shadow testing is the feature to name because a model returns valid-looking responses that are simply wrong, so error rate and latency alone will not catch a bad model. Comparing output distributions between variants is the check that does.

Question 9

What is Amazon Bedrock?

Accepted Answer

Bedrock provides serverless API access to foundation models from Anthropic, Meta, Mistral, Cohere, Amazon and others, with a common interface, enterprise controls — data not used for training, VPC endpoints, KMS encryption, IAM, CloudTrail — plus knowledge bases for RAG, agents, guardrails, model evaluation and fine-tuning. The enterprise-controls point is why an organisation uses Bedrock rather than a consumer API, and it is what an AWS interviewer listens for. Provisioned throughput for guaranteed capacity and predictable cost at high volume is the other feature worth naming.

Question 10

What is retrieval-augmented generation and how would you build it on AWS?

Accepted Answer

RAG grounds a language model in your own data: embed documents into vectors, store them in a vector index, retrieve the most relevant chunks for a query, and include them in the prompt. On AWS use a Bedrock embedding model, a vector store such as OpenSearch Serverless, Aurora with pgvector or S3 Vectors, and a Bedrock model for generation — or Bedrock Knowledge Bases for a managed end-to-end version. The decisions that matter are chunking strategy, embedding model choice and whether you rerank before prompting. Naming evaluation — measuring retrieval recall separately from answer quality — is what distinguishes someone who has shipped RAG from someone who has read about it.

Question 11

What is an embedding and what is the operational catch?

Accepted Answer

An embedding is a dense numeric vector representing meaning, such that semantically similar items are close in vector space, enabling similarity search, recommendation and clustering. The catch is that embeddings from different models are not comparable, so changing the embedding model requires re-embedding the entire corpus. That re-embedding cost is a real operational consideration people discover late, particularly with a large document corpus. Versioning the index alongside the model, so you can switch atomically, is the practical mitigation.

Question 12

What is prompt engineering versus fine-tuning versus RAG?

Accepted Answer

Prompt engineering shapes behaviour with instructions and examples and costs nothing to change. RAG supplies external knowledge at inference time, right when the model needs facts it was not trained on and those facts change. Fine-tuning adjusts weights on your examples, right for teaching a consistent style, format or task behaviour prompting cannot reliably achieve. The rule is that RAG solves knowledge problems and fine-tuning solves behaviour problems, and you try prompting first because iteration is free. Candidates who propose fine-tuning to add facts have the mental model backwards, which is exactly what interviewers probe.

Question 13

What are Bedrock Guardrails?

Accepted Answer

Guardrails apply configurable policies to model inputs and outputs — denied topics, content filters for harmful categories, word filters, PII detection with masking or blocking, and contextual grounding checks that flag responses not supported by the retrieved source. The contextual grounding check is the interesting one for RAG, because it detects hallucination by measuring whether the response is supported by the retrieved context. Applying guardrails independently of the model means they still apply if you switch models.

Question 14

What are Bedrock Agents and tool use?

Accepted Answer

An agent uses a foundation model to plan and execute multi-step tasks, calling defined actions — Lambda functions or APIs — and querying knowledge bases, with the model deciding which tool to call and with what arguments based on the user request. The security point to raise unprompted is that the model chooses the tool call based on untrusted input, so your code must validate arguments and enforce authorisation independently. Treating a model's tool-call request as authorised is a serious vulnerability class.

Question 15

How do you control the cost of generative AI on AWS?

Accepted Answer

Choose the smallest model meeting quality requirements, since lighter models are dramatically cheaper; cap output tokens; trim prompts and avoid resending unchanged context; use prompt caching for repeated large contexts; cache responses for repeated queries; batch where latency permits; and consider provisioned throughput for high steady usage. Prompt caching is the lever most teams miss: a RAG system resending the same large system prompt on every request pays for those input tokens every time. Naming it alongside model right-sizing shows genuine cost engineering rather than generic advice.

Question 16

How do you evaluate a machine learning model before deploying it?

Accepted Answer

Hold out a test set the model never saw; evaluate with metrics matched to the business problem — precision and recall or AUC for imbalanced classification rather than accuracy; check performance across important slices such as region or customer segment; compare against a simple baseline; and validate on the most recent data to catch temporal drift. Slice evaluation is what separates a thoughtful answer: an overall 95% accuracy can hide complete failure for a minority segment. For generative models, evaluation shifts to human review, model-as-judge scoring and task benchmarks, which Bedrock model evaluation supports.

Question 17

What is the difference between precision and recall and when do you optimise for each?

Accepted Answer

Precision is the fraction of positive predictions that are correct; recall is the fraction of actual positives found. Optimise precision when a false positive is costly — blocking a legitimate transaction as fraud. Optimise recall when a false negative is costly — missing a disease diagnosis or a security breach. The framing that scores is that the threshold is a business decision, not a modelling one: the same model gives different precision and recall at different thresholds, so the real question is what each error costs.

Question 18

What is overfitting and how do you detect it?

Accepted Answer

Overfitting is learning noise specific to the training data, so the model performs much worse on unseen data. Detect it as a large gap between training and validation performance. Prevent it with more or more varied data, regularisation, simpler models, early stopping, dropout and proper cross-validation. The complication worth naming is that a validation set contaminated by leakage hides overfitting entirely, which is the failure behind many models that look excellent and fail in production. SageMaker automatic model tuning can optimise regularisation strength directly.

Question 19

What is SageMaker automatic model tuning?

Accepted Answer

Automatic model tuning runs many training jobs with different hyperparameter values, using Bayesian optimisation or Hyperband to choose promising combinations rather than exhaustive search, and reports the best job by the objective metric, with configurable parallelism and early stopping. Bayesian search reaches a good configuration in far fewer trials than grid or random search, which matters because each trial costs GPU time. Higher parallelism is faster but gives the optimiser less information per round, which is the trade-off to name.

Question 20

What AWS AI services exist beyond SageMaker and Bedrock?

Accepted Answer

Managed AI services requiring no model building: Rekognition for images and video, Textract for document extraction, Comprehend for natural language, Transcribe for speech to text, Polly for text to speech, Translate, Personalize for recommendations, Forecast, Fraud Detector and Kendra for enterprise search. The guidance is to check whether a managed service solves the problem before building a model, because it is faster, cheaper and needs no ML expertise. Naming Textract for document processing rather than training a custom model is the kind of pragmatic answer interviewers reward.

Question 21

How do you handle sensitive data in an ML pipeline?

Accepted Answer

Discover and classify with Macie; de-identify at ingestion through tokenisation or masking; apply column and row level access with Lake Formation; encrypt with customer managed KMS keys; keep the pipeline inside a VPC with endpoints and no internet access; and ensure inference request logging does not capture raw sensitive payloads. Inference logging is the trap: enabling data capture for monitoring can silently persist sensitive inputs in a bucket with weaker controls than the source. Naming that specific leak is a strong signal of practical security awareness.

Question 22

What is SageMaker Clarify?

Accepted Answer

Clarify detects bias in data and models — before training, after training and in deployed endpoints — across configurable facets, and provides feature-importance explanations using SHAP so you can see which inputs drove a prediction. The caution is that explanations describe the model, not the world: a high attribution on a proxy variable reveals what the model uses, not causation. Overstating attributions as causal is a common and consequential mistake in regulated decisions.

Question 23

When would you retrain a model and how do you decide?

Accepted Answer

On a schedule matched to how fast the domain changes; on a drift or data-quality alert from Model Monitor; or on a measured performance drop once ground-truth labels arrive. The trigger should be automatic but the deployment gated — a retrained model is promoted only if it beats the incumbent on the evaluation set. The gate is essential: automatic retraining without one can deploy a model trained on corrupted or drifted data, making things worse. Champion-challenger evaluation before promotion is the mechanism to name.

Question 24

How do you serve a model with low latency and controlled cost?

Accepted Answer

Size the endpoint to traffic with autoscaling and a sensible minimum; use an accelerator only if the model needs it; consider Inferentia for cost-efficient inference; batch requests where the client tolerates it; cache repeated predictions; use multi-model endpoints to share infrastructure across many small models; and consider a smaller distilled model. Distillation is the lever candidates least often mention and frequently the largest saving — serving a model a tenth the size at nearly the same accuracy beats any infrastructure tuning. Multi-model endpoints are the AWS-specific mechanism for many low-traffic models.

Question 25

What are Inferentia and Trainium?

Accepted Answer

AWS-designed accelerators: Trainium for training and Inferentia for inference, offering better price-performance than comparable GPU instances for supported models, accessed through the Neuron SDK which compiles models for the hardware. The constraint is model and operator support through Neuron — not every architecture compiles cleanly, so validation is required before committing. The saving is meaningful enough that it is worth testing for any high-volume inference workload.

Question 26

What is a multi-model endpoint?

Accepted Answer

A multi-model endpoint hosts many models behind one endpoint on shared infrastructure, loading each model from S3 into memory on demand and evicting unused ones, so hundreds of low-traffic models share a small fleet instead of each needing its own endpoint. The use case is per-customer or per-region models where each individually has little traffic. The trade-off is a cold-load latency when a model is not resident, so it suits workloads tolerant of occasional slower first requests.

Question 27

How would you build a recommendation system on AWS?

Accepted Answer

For a managed route, Amazon Personalize handles data ingestion, model training and real-time recommendations. For a custom route: build user and item embeddings, index item embeddings in a vector store for candidate retrieval, then rank candidates with a model including context features, serving features from Feature Store and the model from a SageMaker endpoint. The two-stage retrieval-then-ranking architecture is the substance — you cannot score millions of items per request, so you retrieve a few hundred cheaply and rank them expensively. Naming that structure demonstrates real recommender knowledge.

Question 28

What is responsible AI and what does AWS provide?

Accepted Answer

Responsible AI covers fairness, explainability, safety, privacy and governance. AWS provides Clarify for bias detection and explanations, Bedrock Guardrails for content and topic policies, model cards for documentation, and Bedrock model evaluation. The organisational half — review processes and defined acceptable use — is not a product. The strongest answer acknowledges that tooling supports but does not deliver responsible AI: deciding what fairness means for a specific decision is a human judgement. Interviewers for senior roles look for that distinction rather than a product list.

Question 29

What is Amazon SageMaker Ground Truth?

Accepted Answer

Ground Truth manages data labelling with human workforces — your own team, a vendor or Mechanical Turk — with built-in workflows for common tasks, active learning to reduce labelling volume by auto-labelling confident examples, and consolidation across multiple annotators. Label quality is usually the binding constraint on supervised model quality, and inter-annotator agreement is the metric to track. Active learning reducing the number of examples a human must label is the cost lever worth naming.

Question 30

How do you handle a model that performs well in evaluation but poorly in production?

Accepted Answer

Check for training-serving skew in feature computation; compare live feature distributions against the training baseline; look for target leakage in the training data; verify serving preprocessing matches training exactly; and check whether the production population differs from the training population in a way the evaluation split hid. The ordering matters: skew and leakage explain the great majority of such cases, and both are data problems rather than model problems. A candidate who immediately proposes a different architecture is looking in the wrong place, which is what the question tests.

Question 31

What is the difference between supervised, unsupervised and reinforcement learning?

Accepted Answer

Supervised learning trains on labelled examples to predict a target — classification and regression. Unsupervised learning finds structure in unlabelled data — clustering, dimensionality reduction, anomaly detection. Reinforcement learning learns a policy by acting in an environment and receiving rewards. Grounding each in an AWS service — SageMaker built-in algorithms for supervised and unsupervised, custom training for reinforcement learning — turns theory into platform knowledge. Naming anomaly detection as the unsupervised case most teams actually use is a practical touch.

Question 32

What is feature engineering and where does it happen on AWS?

Accepted Answer

Feature engineering transforms raw data into model inputs — aggregations, ratios, time windows, encodings, normalisation. On AWS it happens in Athena or Redshift SQL for tabular data, Glue or SageMaker Processing for batch, Managed Flink for streaming features, with results stored in Feature Store for consistent reuse. The rule to state is that whatever computes a feature for training must also compute it for serving, which is exactly why the feature store exists. Framing feature engineering as a consistency problem, not only a modelling one, connects it to production reality.

Question 33

How do you secure a SageMaker environment?

Accepted Answer

Run notebooks and training in a VPC with no direct internet access, using VPC endpoints for S3 and SageMaker APIs; encrypt volumes and artefacts with customer managed KMS keys; use IAM roles per project rather than a shared broad role; enable network isolation on training jobs; and log everything to CloudTrail. Network isolation on training jobs is the specific control worth naming: it prevents the training container making any network calls at all, which matters when running third-party code. Notebook instances with internet access are a common and easily-closed exposure.

Question 34

What is a vector database and which options exist on AWS?

Accepted Answer

A vector database stores embeddings and performs approximate nearest neighbour search efficiently. On AWS the options are OpenSearch Service and OpenSearch Serverless with k-NN, Aurora PostgreSQL with pgvector, DocumentDB vector search, MemoryDB, Neptune Analytics and S3 Vectors for cost-efficient large-scale storage. The trade-off inherent in approximate search is recall versus latency and cost — an ANN index does not guarantee the true nearest neighbours, and tuning parameters control that balance. Acknowledging approximation rather than describing it as exact is the correctness marker.

Question 35

How do you evaluate a RAG system?

Accepted Answer

Evaluate retrieval and generation separately: retrieval by recall and precision at k against a labelled set of question-to-document pairs, and generation by faithfulness to the retrieved context, answer relevance and correctness, using human review or a model-as-judge with a rubric. Track them independently because they fail differently. Separating the two is the key insight: a wrong answer may be a retrieval failure — the right document was never fetched — or a generation failure. Conflating them means you optimise the wrong half, which is the most common RAG debugging error.

Question 36

What is prompt injection and how do you mitigate it?

Accepted Answer

Prompt injection is untrusted content — a user message, a retrieved document, a web page — containing instructions the model follows, overriding your intent. Mitigate by treating all retrieved and user content as data rather than instructions, validating and authorising every tool call independently in code, applying guardrails, and limiting what the model can do. The critical principle is that the model's output must never be the authorisation decision: if an agent can delete a record, your code must check the caller's permission, not trust that the model decided it was appropriate. That is the mitigation that actually holds.

Question 37

Design an end-to-end ML platform on AWS for a team of ten data scientists.

Accepted Answer

Governed data in S3 and Redshift with Lake Formation column-level control. Features defined once in SageMaker Feature Store. Experimentation in SageMaker Studio inside a VPC with no internet, per-project IAM roles and idle shutdown. Training as version-controlled SageMaker Pipelines triggered by CodePipeline, writing to the Model Registry with lineage. Promotion gated on automated evaluation including slice metrics. Deployment via deployment guardrails with shadow testing, or batch transform where interactive latency is not needed. Model Monitor watching skew and drift and triggering retraining. Everything provisioned with IaC, encrypted with customer managed keys, and cost-attributed by project tag. The closing scenario. The senior markers are the evaluation gate before promotion, the feature store solving skew rather than being a convenience, shadow testing because model failures are silent, and treating monitoring as a retraining trigger rather than a dashboard.

AWS SageMaker, Bedrock & AI/ML Interview Questions and Answers

How to use this set