Hard Difficulty Questions

k-anonymity

Why wrong: k-anonymity generalizes data to make individuals indistinguishable, but does not add noise during training.

Federated learning

Why wrong: Federated learning trains across decentralized data but requires differential privacy for formal guarantees.

Homomorphic encryption

Why wrong: Homomorphic encryption allows computation on encrypted data, not noise injection for privacy during training.

Enable public endpoint with TLS and restrict access using a VPN

Why wrong: Public endpoint still leaves the endpoint internet-facing; a VPN is additional but less secure than VPC isolation.

Use SageMaker Model Monitor to detect anomalies in request data

Why wrong: Model Monitor monitors data quality, not encryption or network access.

Create a SageMaker endpoint in a private subnet without an internet gateway, and use AWS PrivateLink

Why wrong: PrivateLink is correct but the endpoint itself can be accessed via PrivateLink without HTTPS; you still need TLS.

SageMaker Model Monitor

Why wrong: Monitors for data and model quality drift.

SageMaker Autopilot

Why wrong: Automates ML model creation.

SageMaker Debugger

Why wrong: Helps debug training problems.

Configure a Bedrock Guardrail to log all invocations to a CloudWatch Logs group in the required region

Why wrong: Guardrails do not log full invocation content; they log only guardrail evaluations.

Use AWS CloudTrail to log all API calls and store them in the required region

Why wrong: CloudTrail logs API calls, not model invocation content (prompts/responses). The requirement is for inference logs.

Use AWS Global Accelerator to route traffic to the required region

Why wrong: Global Accelerator improves performance but does not control log storage location.

Amazon Comprehend

Why wrong: Comprehend can analyze sentiment but is not designed for real-time content filtering in Bedrock.

Amazon Rekognition

Why wrong: Amazon Rekognition analyses images and videos for objects, scenes, and faces, but lacks the natural language processing capability to scan text for offensive language, which is the requirement in this scenario. It is tempting because Rekognition includes content moderation for unsafe visual content, such as explicit imagery, making it a plausible choice for filtering inappropriate material in a different context—specifically, when moderating image or video assets rather than generated text.

AWS WAF

Why wrong: WAF protects web applications from attacks, not content moderation.

Always use the largest available model for best results

Why wrong: Using the largest model increases cost without guaranteed throughput benefit.

Fine-tune a large foundation model on the company's codebase

Why wrong: Fine-tuning incurs additional training and hosting costs, not a cost-saving measure.

Fine-tune the model using SageMaker Ground Truth and increase the training epochs.

Why wrong: Fine-tuning addresses style but not immediate prompt adherence; Ground Truth is for labeling.

Increase the max token count and use a larger model variant.

Why wrong: Max tokens apply to text generation, not image generation.

Use a different foundation model and increase the image resolution.

Why wrong: Model switch might help, but resolution alone doesn't fix artifacts.

Amazon CloudFront

Why wrong: CloudFront caches content at edge locations, not model outputs.

Amazon DynamoDB

Why wrong: DynamoDB can be used but is not as fast as a dedicated cache for this use case.

Amazon S3

Why wrong: S3 has higher latency and is not designed as a cache.

Deploy on a SageMaker multi-model endpoint

Why wrong: Multi-model endpoint is designed for multiple models; for a single model, it adds unnecessary complexity.

Use batch transform

Why wrong: Batch transform is for asynchronous, not real-time.

Use a single large instance type

Why wrong: Overprovisioning leads to high cost when traffic is low.

Fine-tune the model on de-identified data only

Why wrong: Fine-tuning is expensive and time-consuming; it does not guarantee the model will never output PII.

Use a prompt engineering technique to instruct the model to avoid PII

Why wrong: Prompt engineering may not be sufficient as models can ignore instructions or be manipulated.

Post-process the output with a regex filter

Why wrong: Regex filters are prone to errors and can miss novel patterns of PII.

The action should be "bedrock:InvokeModelWithResponseStream".

Why wrong: InvokeModel is the correct action for synchronous invocation; InvokeModelWithResponseStream is for streaming.

The resource ARN is missing the account ID.

Why wrong: The account ID is present (123456789012).

The statement is missing a condition for the model ID.

Why wrong: Conditions are optional; the resource ARN already specifies the model ID.

The API automatically reserves 2,000 tokens for output, reducing available input capacity

Why wrong: There is no automatic reservation; the total must fit within the window.

The model requires a minimum of 512 tokens for internal processing

Why wrong: No such minimum is required.

The model's context window counts only input tokens; output tokens are separate

Why wrong: Context window includes both input and output tokens.

Use Amazon Comprehend to analyze customer reviews and correlate with sales

Why wrong: Comprehend is for NLP, not time-series forecasting.

Upload data to Amazon QuickSight and use its built-in forecasting widget

Why wrong: QuickSight's forecasting is limited for many SKUs and lacks advanced features.

Train individual ARIMA models for each SKU using Amazon SageMaker built-in algorithms

Why wrong: Manual per-SKU modeling is labor-intensive and ARIMA may not capture external factors well.

Fine-tune the model on a large corpus of customer service transcripts to improve domain knowledge.

Why wrong: Fine-tuning is costly and may not be necessary if RAG can provide context.

Select a smaller, faster model that trades some accuracy for throughput.

Why wrong: This may compromise accuracy, which is critical for customer service.

Increase the model's maximum token limit to handle longer customer queries.

Why wrong: Higher token limits increase cost and latency, not a recommended optimization.

Amazon Transcribe and Amazon Polly

Why wrong: Transcribe converts speech to text, Polly converts text to speech — neither extracts text from documents.

Amazon Kendra and Amazon Lex

Why wrong: Kendra is an enterprise search service; Lex is for conversational interfaces.

Amazon Rekognition and Amazon Comprehend

Why wrong: Rekognition is for image analysis, not document text extraction; Comprehend is for NLP but doesn't extract text from images.

Increase the number of epochs

Why wrong: More epochs often lead to overfitting.

Use a smaller learning rate

Why wrong: Smaller learning rate may not prevent overfitting directly.

The StoppingCondition is too short

Why wrong: Would cause timeout, not resource not found.

The instance type ml.m5.large is not supported

Why wrong: ml.m5.large is supported for bias jobs.

The IAM role lacks permissions to DescribeBaselineJob

Why wrong: Would cause access denied, not resource not found.

Amazon SageMaker

Why wrong: SageMaker is for model training and deployment, not a core RAG component.

AWS Lambda

Why wrong: Lambda can be used in action groups for agents, but not as a primary RAG component.

Amazon Bedrock Guardrails

Why wrong: Guardrails are optional safety features, not part of the core RAG pipeline.

The LLM used for generation has too few parameters

Why wrong: Model size affects generation quality but not retrieval relevance.

The agent's action group has misconfigured Lambda functions

Why wrong: Action groups are used for tool calling, not for the core RAG retrieval pipeline.

The vector store index is not encrypted at rest

Why wrong: Encryption does not affect retrieval accuracy.