TechNuggets Academy

Foundation Model Integration, Data Management, and Compliance

Free AWS Certified GenAI Developer (AIP-C01) practice — 6 questions on Foundation Model Integration, Data Management, and Compliance, with explanations. No sign-up. Full 12-question mixed test →

Question 1 of 6 · Domain 1: Foundation Model Integration, Data Management, and Compliance
A financial services company is building an AI application that processes customer loan applications. The application must ensure that customer PII is never sent to the foundation model provider's servers and must maintain an audit trail of all model interactions for 7 years to meet regulatory requirements. The company wants to use Amazon Bedrock with Claude models. Which combination of features should the architect implement?
This solution correctly addresses all requirements: Model invocation logging to S3 captures full request/response payloads for audit trails with appropriate retention, AWS PrivateLink ensures traffic never traverses the public internet (data stays within AWS network but note that Bedrock processes data in AWS-managed infrastructure, not customer VPC), and field-level tokenization removes PII before it reaches the model, ensuring sensitive data never leaves customer control. S3 supports long-term retention policies required for compliance.
Question 2 of 6 · Domain 1: Foundation Model Integration, Data Management, and Compliance
A healthcare technology company needs to fine-tune Amazon Bedrock's Titan Text model on 50,000 proprietary clinical notes containing patient diagnoses and treatment protocols. The data contains HIPAA-regulated PHI. The company's compliance team requires that: (1) the training data must remain encrypted with keys they control, (2) the fine-tuned model must not be accessible to AWS personnel, and (3) they must be able to prove the training data was deleted after fine-tuning. Which approach meets these requirements?
This is the correct approach for HIPAA-compliant fine-tuning: SSE-KMS with customer-managed keys ensures customer controls encryption keys; Bedrock fine-tuning keeps the resulting custom model within the customer's account (AWS personnel cannot access it); the custom model is account-specific and private; deleting S3 objects with versioning disabled ensures permanent deletion provable through CloudTrail logs and S3 access logs; Bedrock supports fine-tuning Titan models with customer data that remains in customer's account/region.
Question 3 of 6 · Domain 1: Foundation Model Integration, Data Management, and Compliance
An AI application uses Amazon Bedrock with the Anthropic Claude 3 Sonnet model to analyze customer support tickets. During testing, the team discovers that 3% of responses contain hallucinated product feature names that don't exist in their catalog. The application architecture currently uses a simple invoke_model API call with default parameters. Which solution MOST effectively reduces hallucinations while maintaining response quality?
RAG with Bedrock Knowledge Bases is the most effective solution because it grounds the model's responses in verified source documents (product documentation). The retrieve_and_generate API automatically retrieves relevant context from the knowledge base and includes it in the prompt, giving the model factual information to work with. This architecturally prevents hallucinations by providing verified facts rather than relying on the model's training data. This is the recommended AWS pattern for reducing hallucinations in domain-specific applications.
Question 4 of 6 · Domain 1: Foundation Model Integration, Data Management, and Compliance
A global e-commerce company is deploying an AI-powered product recommendation system using Amazon Bedrock across multiple AWS regions (us-east-1, eu-west-1, ap-southeast-1). Due to GDPR and data residency requirements, European customer data must be processed only in eu-west-1, and the model responses must be identical across regions for A/B testing validity. The current architecture uses Amazon Bedrock on-demand inference with the Titan Embeddings model. Which configuration ensures compliance while maintaining consistent model behavior?
This solution correctly addresses all requirements: Using on-demand inference with identical model IDs (e.g., amazon.titan-embed-text-v1) and parameters ensures consistent model behavior across regions since Bedrock maintains model version consistency; application-level region routing ensures customer data is processed in the appropriate region for compliance; DynamoDB Global Tables enables fast customer-region lookups while maintaining eu-west-1 as the authoritative region for EU customer data. This architecture respects data residency without complex cross-region data transfers.
Question 5 of 6 · Domain 1: Foundation Model Integration, Data Management, and Compliance
A legal tech company is building a contract analysis system using Amazon Bedrock with Claude 3 Opus. The system must process contracts containing sensitive merger & acquisition data. The security team mandates: (1) all prompt data must be encrypted in transit and at rest using FIPS 140-2 validated cryptographic modules, (2) model invocation logs must be immutable for 10 years, and (3) they must be able to immediately revoke all access to the AI system if a security breach is detected. The system processes 10,000 contracts daily. Which architecture meets these requirements while optimizing for cost?
This solution correctly meets all requirements: FIPS endpoints satisfy FIPS 140-2 requirement for cryptographic modules; S3 Object Lock in compliance mode creates truly immutable logs for 10 years (cannot be deleted even by root); Glacier Deep Archive after 90 days optimizes cost for long-term retention; AWS STS temporary credentials combined with SCPs enable immediate revocation by denying at the organization level; customer-managed KMS keys provide additional control. This is cost-optimized through Glacier transition and on-demand inference for variable workload.
Question 6 of 6 · Domain 1: Foundation Model Integration, Data Management, and Compliance
A media company uses Amazon Bedrock with the Stability AI SDXL model to generate images for marketing campaigns. They discover that 12% of generated images inadvertently contain competitor brand logos or copyrighted characters. The company needs to prevent this while maintaining creative output quality and meeting a 2-second response time SLA. The current implementation uses invoke_model with default parameters and generates 50,000 images per day. Which solution BEST addresses this requirement?
This is the most effective solution: Negative prompts directly instruct the image generation model what NOT to include, which is the primary mechanism for controlling Stable Diffusion output; Amazon Rekognition Content Moderation can detect logos and inappropriate content in images post-generation; Step Functions provides orchestration for retry logic without adding latency to successful requests; this combines prevention (negative prompt) with detection (Rekognition) and remediation (retry). The approach addresses the root cause while maintaining the 2-second SLA for clean images (only retries add latency).
Ready for the real thing?

The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.

$54.99 $17.99 with code FREETEST33 — valid through August 22.

Get my $17.99 deal →