Question 1

An insurance company has deployed a Generative AI chatbot using Amazon Bedrock to explain policy details to customers. To comply with strict financial regulations, the solution requires a continuous monitoring and automated remediation system for two specific governance issues:Misuse Remediation: If a specific user triggers the "Hate Speech" content filter in Amazon Bedrock Guardrails more than 5 times in a single minute, their access must be automatically and immediately suspended.Bias Monitoring: The system must strictly monitor the model's responses over time to detect if the terminology becomes statistically biased against specific demographics compared to a pre-established "Golden Set" baseline.Which combination of mechanisms satisfies these requirements?

Options :

A : Implement a scheduled AWS Step Functions workflow that uses Amazon SageMaker Clarify to analyze the aggregated inference logs stored in Amazon S3. Configure the job to calculate bias metrics against the baseline and publish findings to CloudWatch.

B : Deploy Amazon SageMaker Model Monitor on the Amazon Bedrock endpoint. Enable the "Real-time Concept Drift" feature to block any inference response that deviates more than 5% from the training data distribution.

C : Configure an Amazon EventBridge Rule to match the aws.bedrock.GuardrailIntervention event pattern. Set the target as an AWS Systems Manager Run Command to disable the user's account in the identity provider.

D : Enable Amazon Macie on the bucket containing the inference logs. configure a custom data identifier for "Biased Terminology" and set up an automated remediation job to delete non-compliant logs.

E : Create an Amazon CloudWatch Alarm that monitors the GuardrailInterventionCount metric with a distinct dimension for the user ID. Configure the alarm action to trigger an AWS Lambda function that revokes the user's IAM permissions or API keys.

Answer: A,E

Question 2

A financial services company has 15 development teams building different microservices using Amazon Bedrock. Code reviews reveal inconsistent error handling patterns, varying retry logic implementations, and different approaches to streaming responses across services. The architecture team defined standard Bedrock integration patterns, but manual enforcement through code reviews is slow and inconsistent. The company needs a solution that automatically identifies non-standard implementations and helps developers refactor code to match approved patterns across the entire multi-repository codebase.Which solution enforces consistent integration patterns MOST effectively?

Options :

A : Create an AWS CodeCommit repository with approved code templates and integration examples. Distribute the repository link to all teams with instructions to reference it when implementing Bedrock features.

B : Implement pre-commit hooks in Git that run custom Python scripts to validate Bedrock API calls against a JSON schema defining allowed patterns before allowing commits.

C : Deploy SonarQube on Amazon EC2 to perform static code analysis. Configure custom rules that flag non-standard Bedrock integration patterns and generate reports for teams to manually fix.

D : Configure Amazon Q Developer to scan the codebase, identify Bedrock integration anti-patterns, generate refactored code following company standards, and apply the consistent changes across all microservice repositories.

Answer: D

Question 3

A large financial services enterprise is building a centralized platform to allow internal development teams to consume Foundation Models (FMs) from Amazon Bedrock. To ensure strict compliance and governance, the organization requires a "GenAI Gateway" architecture.The architecture must meet the following requirements:Prevent "Shadow AI" by forcing all internal applications to route requests through a single control point that logs every prompt and response for auditing.Automate the testing of prompt templates. Whenever a prompt engineer updates a standard system prompt, it must be automatically evaluated for toxicity and accuracy before being made available to applications.Ensure that PII is automatically redacted from prompts before they are sent to the model provider.Which combination of actions should you implement to achieve this?

Options :

A : Grant all internal application roles direct IAM permissions to bedrock:InvokeModel. Enable AWS CloudTrail data events for Amazon Bedrock to capture the request payloads and response bodies for auditing purposes.

B : Use Amazon Inspector within the CI/CD pipeline to scan the weights of the Amazon Bedrock base models for vulnerabilities and malware before approving them for use in the production environment.

C : Implement AWS AppSync with a direct data source to Amazon Bedrock. Use VTL resolvers to hardcode the system prompts within the API definition, ensuring that developers cannot modify them without redeploying the API.

D : Develop a centralized REST API using Amazon API Gateway and AWS Lambda. Configure the Lambda function to act as an abstraction layer that intercepts requests, uses Amazon Comprehend to detect and redact PII, logs the transaction to Amazon CloudWatch Logs, and then invokes Amazon Bedrock.

E : Create an AWS CodePipeline for prompt versioning. Include an AWS CodeBuild stage that runs an automated evaluation script (using a library like Ragas or specific Bedrock evaluations) against the new prompt candidate. Configure the pipeline to fail and rollback if the toxicity score exceeds a defined threshold.

Answer: D,E

Question 4

A marketing analytics company generates monthly reports that include AI-generated summaries of customer feedback. They have 10,000 customer feedback entries to summarize each month. The summaries are not time-sensitive and can be processed overnight. The team wants to optimize costs compared to processing requests one-by-one in real-time.Which processing approach should the developer use to reduce costs?

Options :

A : Invoke the foundation model in real-time for each customer feedback entry as soon as it arrives

B : Use provisioned throughput to guarantee capacity for all 10,000 requests

C : Use batch processing to send multiple requests together and process them during off-peak hours

D : Enable auto-scaling to handle the large burst of requests

Answer: C

Question 5

A legal tech company uses Amazon Bedrock to extract specific liability clauses from uploaded contracts. Users have reported instances where the model "hallucinates" clauses that do not exist in the source text.Which mechanisms should the developer implement to verify accuracy and reduce these hallucinations?

Options :

A : Define a JSON Schema representing the expected liability data structure and include it in the model instructions to enforce constrained, deterministic output.

B : Implement a post-processing workflow that calculates the semantic similarity score between the generated response and the source text, rejecting responses that fall below a defined confidence threshold.

C : Configure Amazon Bedrock Guardrails with a "Sensitive Information" filter to redact names and addresses found in the contract clauses.

D : Increase the inference temperature parameter to 0.9 and enable Top-P sampling to broaden the diversity of the generated clauses.

E : Use Amazon CloudWatch to monitor the InvocationLatency and ModelTokenCount metrics to identify anomalies in the model's processing speed.

Answer: A,B