AI Recommendation Architecture with AgentCore and Immutable Audit
A financial company is building an AI-powered recommendation system. The system uses real-time market data and verified internal documents. The system must include evaluation capabili-ties that collect explicit user ratings and detect problematic responses in real time. The system must include evaluation capabilities that track model performance over time and support A/B testing of different prompt or model variants. The system must orchestrate multi-step workflows. The workflows must retrieve trusted knowledge, validate inputs, and generate recommendations. The system must securely store all user feedback, flagged events, and audit logs in a compliance-ready immutable repository. The repository must be write- once, tamper-evident storage that is suitable for regulated audit reten-tion. Which architecture will meet these requirements?
Community Insight
The exam tests whether you can assemble an architecture that meets ALL stated requirements simultaneously, and the common trap is picking a design that has orchestration or Object Lock but quietly fails on real-time detection, continuous monitoring, or A/B testing.
This question tests which AWS architecture satisfies real-time evaluation, multi-step agentic orchestration, A/B testing, and compliance-ready immutable audit storage for a financial recommendation system. The correct architecture (C) chains Amazon Bedrock AgentCore with curated Kendra retrieval, Lambda-based input and post-generation validation, traffic-split A/B testing, S3 Object Lock for tamper-evident logs, and SageMaker Model Monitor for continuous drift tracking.
The most tempting wrong answers are A and D. Option A uses scheduled Athena queries on CloudTrail logs, which is a batch job and cannot detect problematic responses in real time, and it omits A/B testing and continuous model performance tracking. Option D relies on quarterly manual audits and Amazon RDS instead of write-once, tamper-evident storage, directly violating the immutable repository requirement.
Community Discussion (1 comments)
Comments & Corrections
No comments yet — spotted an error or have a note? Share it below.
Expert Analysis
Why the Answer Is Correct
Option C is the only design that satisfies every clause of the requirements in a single coherent architecture. Amazon Bedrock AgentCore orchestrates the multi-step workflow, retrieving trusted knowledge from curated Amazon Kendra indexes (verified internal documents), while an AWS Lambda function sanitizes input before it reaches the model — satisfying input validation. A second function performs post-generation fact-consistency checking, which delivers real-time detection of problematic responses, and a third function splits traffic to enable true A/B testing of prompt and model variants. User feedback (explicit ratings) lands in DynamoDB, audit logs flow through CloudTrail into Amazon S3 with S3 Object Lock enabled for write-once, tamper-evident regulated retention, and SageMaker Model Monitor continuously tracks data drift and output anomalies to satisfy the "track model performance over time" requirement.Why the Other Options Are Wrong
Option A has orchestration via AgentCore and even Object Lock, but its anomaly detection runs as scheduled Amazon Athena queries against CloudTrail logs — this is periodic batch analysis, not the real-time problematic-response detection the question demands; it also never establishes A/B testing or continuous model performance monitoring. Option B is a thin, single-hop pipeline (API Gateway invoking a Bedrock FM with Kendra) that lacks multi-step workflow orchestration, omits input validation and post-generation consistency checks entirely, and misuses Amazon Macie — a sensitive-data discovery service — as a hallucination/problematic-content detector; it also provides no A/B testing. Option D fails the hardest: Amazon Lex aliases routing to SageMaker endpoints does not constitute the required orchestration, quarterly manual audits are the opposite of real-time detection, and storing feedback in Amazon RDS gives mutable storage rather than the mandated write-once, tamper-evident repository.Community Comment Notes
Commenter TNT87 also selected C, observing that orchestration "perfectly handles multi-step conversational and retrieval tasks" and "orchestrates pulling trusted, verified domain data directly from customized Amazon Kendra indices." This aligns with the requirement for retrieving trusted knowledge as part of a multi-step workflow. That said, the Kendra and AgentCore orchestration is only one piece of why C wins — the deciding factors across the other options are real-time detection, continuous Model Monitor drift tracking, and S3 Object Lock immutability.Official Reference
Exam Strategy
When an architecture question lists many mandatory capabilities, build a mental checklist and verify each option against every clause — orchestration, real-time detection, continuous monitoring, A/B testing, and immutable storage. The winning option is almost always the one with no gaps, while distractors fail on exactly one subtle requirement like batch-versus-real-time or mutable-versus-WORM storage.
Frequently Asked Questions
Why isn't option A correct when it also uses AgentCore and S3 Object Lock?
Option A detects anomalies only via scheduled Athena queries on CloudTrail logs, which is batch, not real-time. It also lacks A/B testing and continuous model performance tracking that option C provides.
Why is S3 Object Lock required for the audit repository here?
S3 Object Lock provides write-once, tamper-evident (WORM) storage required for regulated audit retention, ensuring feedback, flagged events, and logs cannot be altered or deleted.
Related Analysis
Practice All AIP-C01 Questions
Access 110 questions with complete answers and detailed explanations.
View Full AIP-C01 Practice Test →