Question 1
At a retail bank, engineers need to deploy a large self-hosted model with significant GPU memory needs. The architecture must avoid unnecessary coupling between components. The initial assessment isolates this issue from unrelated infrastructure changes. Which architecture is best supported by the stated constraints?
- Run a large GPU workload in an API Gateway mapping template
- Size accelerator memory, batching, concurrency and model-loading strategy explicitly
- Treat model deployment as a CPU-only static web hosting task
- Assume endpoint throughput is independent of generated token length
Explanation
Correct Answer
B:Self-hosted LLM serving must account for memory, accelerator use and token generation dynamics. In a retail bank, the engineering team should verify this design against its stated objective to deploy a large self-hosted model with significant GPU memory needs, including the requirement that the architecture must avoid unnecessary coupling between components.
Incorrect Answers
A:The suggested action—run a large GPU workload in an API Gateway mapping template—does not establish the required control for a retail bank to deploy a large self-hosted model with significant GPU memory needs. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a retail bank. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—treat model deployment as a CPU-only static web hosting task—does not establish the required control for a retail bank to deploy a large self-hosted model with significant GPU memory needs. The change is applied outside the component responsible for the requirement identified by a retail bank. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—assume endpoint throughput is independent of generated token length—does not establish the required control for a retail bank to deploy a large self-hosted model with significant GPU memory needs. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a retail bank. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 2
At a shipping brokerage, engineers need to support predictable sustained inference volume with committed capacity. The architecture must avoid unnecessary coupling between components. A recent review found that the current design has no reliable safeguard for this case. Which architecture is best supported by the stated constraints?
- Assume on-demand calls guarantee reserved capacity
- Use S3 storage classes to reserve model tokens
- Increase client HTTP timeout to create capacity
- Evaluate Bedrock Provisioned Throughput eligibility and economics for the selected model
Explanation
Correct Answer
D:Provisioned throughput can provide a capacity model when supported and properly sized. In a shipping brokerage, the engineering team should verify this design against its stated objective to support predictable sustained inference volume with committed capacity, including the requirement that the architecture must avoid unnecessary coupling between components.
Incorrect Answers
A:The suggested action—assume on-demand calls guarantee reserved capacity—does not establish the required control for a shipping brokerage to support predictable sustained inference volume with committed capacity. It replaces a verifiable engineering control with an untested shortcut in the a shipping brokerage deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
B:The suggested action—use S3 storage classes to reserve model tokens—does not establish the required control for a shipping brokerage to support predictable sustained inference volume with committed capacity. The change is applied outside the component responsible for the requirement identified by a shipping brokerage. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—increase client HTTP timeout to create capacity—does not establish the required control for a shipping brokerage to support predictable sustained inference volume with committed capacity. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a shipping brokerage. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 3
At a security incident response team, engineers need to switch model providers without modifying application business logic. The next release must be evaluated against the previous baseline. The initial assessment isolates this issue from unrelated infrastructure changes. Which implementation most directly satisfies the requirement?
- Introduce a model adapter and externalized routing configuration
- Embed provider-specific request formats in every user interface
- Hard-code one model ID in all business services
- Depend on prompt wording to silently change cloud providers
Explanation
Correct Answer
A:An abstraction layer isolates provider requests and lets configuration drive changes. In a security incident response team, the engineering team should verify this design against its stated objective to switch model providers without modifying application business logic, including the requirement that the next release must be evaluated against the previous baseline.
Incorrect Answers
B:The suggested action—embed provider-specific request formats in every user interface—does not establish the required control for a security incident response team to switch model providers without modifying application business logic. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a security incident response team. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—hard-code one model ID in all business services—does not establish the required control for a security incident response team to switch model providers without modifying application business logic. It replaces a verifiable engineering control with an untested shortcut in the a security incident response team deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—depend on prompt wording to silently change cloud providers—does not establish the required control for a security incident response team to switch model providers without modifying application business logic. The change is applied outside the component responsible for the requirement identified by a security incident response team. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 4
At a commercial property platform, engineers need to debug malformed tool arguments emitted by a model. The implementation must remain maintainable across model upgrades. The initial assessment isolates this issue from unrelated infrastructure changes. Which change should the team prioritize?
- Validate against the tool schema, inspect failed payloads and refine tool definitions
- Execute malformed payloads with administrator rights
- Increase vector search shards without inspecting API errors
- Ignore schema failures until the model guesses the right shape
Explanation
Correct Answer
A:Tool-call diagnosis examines model outputs against executable interface contracts. In a commercial property platform, the engineering team should verify this design against its stated objective to debug malformed tool arguments emitted by a model, including the requirement that the implementation must remain maintainable across model upgrades.
Incorrect Answers
B:The suggested action—execute malformed payloads with administrator rights—does not establish the required control for a commercial property platform to debug malformed tool arguments emitted by a model. The change is applied outside the component responsible for the requirement identified by a commercial property platform. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—increase vector search shards without inspecting API errors—does not establish the required control for a commercial property platform to debug malformed tool arguments emitted by a model. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a commercial property platform. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—ignore schema failures until the model guesses the right shape—does not establish the required control for a commercial property platform to debug malformed tool arguments emitted by a model. It replaces a verifiable engineering control with an untested shortcut in the a commercial property platform deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 5
At a telecom support operation, engineers need to serve a customized model with requirements not supported by the managed model catalog. A controlled rollout is required before broad user exposure. The initial assessment isolates this issue from unrelated infrastructure changes. Which remediation addresses the root problem?
- Assume every custom model is automatically a Bedrock on-demand model
- Use API Gateway as the model inference engine
- Place model weights in a Lambda environment variable
- Evaluate a SageMaker AI inference endpoint and its resource sizing
Explanation
Correct Answer
D:Managed custom inference endpoints support deployment controls for compatible specialized models. In a telecom support operation, the engineering team should verify this design against its stated objective to serve a customized model with requirements not supported by the managed model catalog, including the requirement that a controlled rollout is required before broad user exposure.
Incorrect Answers
A:The suggested action—assume every custom model is automatically a Bedrock on-demand model—does not establish the required control for a telecom support operation to serve a customized model with requirements not supported by the managed model catalog. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a telecom support operation. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
B:The suggested action—use API Gateway as the model inference engine—does not establish the required control for a telecom support operation to serve a customized model with requirements not supported by the managed model catalog. It replaces a verifiable engineering control with an untested shortcut in the a telecom support operation deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—place model weights in a Lambda environment variable—does not establish the required control for a telecom support operation to serve a customized model with requirements not supported by the managed model catalog. The change is applied outside the component responsible for the requirement identified by a telecom support operation. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 6
At an airline operations unit, engineers need to support predictable sustained inference volume with committed capacity. The team must preserve existing identity and authorization boundaries. An architecture review has rejected fixes that only conceal the symptoms. Which remediation addresses the root problem?
- Increase client HTTP timeout to create capacity
- Evaluate Bedrock Provisioned Throughput eligibility and economics for the selected model
- Assume on-demand calls guarantee reserved capacity
- Use S3 storage classes to reserve model tokens
Explanation
Correct Answer
B:Provisioned throughput can provide a capacity model when supported and properly sized. In an airline operations unit, the engineering team should verify this design against its stated objective to support predictable sustained inference volume with committed capacity, including the requirement that the team must preserve existing identity and authorization boundaries.
Incorrect Answers
A:The suggested action—increase client HTTP timeout to create capacity—does not establish the required control for an airline operations unit to support predictable sustained inference volume with committed capacity. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by an airline operations unit. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—assume on-demand calls guarantee reserved capacity—does not establish the required control for an airline operations unit to support predictable sustained inference volume with committed capacity. The change is applied outside the component responsible for the requirement identified by an airline operations unit. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—use S3 storage classes to reserve model tokens—does not establish the required control for an airline operations unit to support predictable sustained inference volume with committed capacity. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by an airline operations unit. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 7
At a nonprofit grant portal, engineers need to decouple customer event ingestion from slow downstream GenAI processing. The team must preserve existing identity and authorization boundaries. The initial assessment isolates this issue from unrelated infrastructure changes. What should be implemented before scaling this workflow?
- Store event payloads in a prompt template as the message queue
- Retry in the user browser after it closes
- Publish an event and process asynchronously with durable retry or queue semantics
- Block the transaction until every AI job finishes
Explanation
Correct Answer
C:Event-driven integration separates business transaction latency from model processing. In a nonprofit grant portal, the engineering team should verify this design against its stated objective to decouple customer event ingestion from slow downstream GenAI processing, including the requirement that the team must preserve existing identity and authorization boundaries.
Incorrect Answers
A:The suggested action—store event payloads in a prompt template as the message queue—does not establish the required control for a nonprofit grant portal to decouple customer event ingestion from slow downstream GenAI processing. The change is applied outside the component responsible for the requirement identified by a nonprofit grant portal. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
B:The suggested action—retry in the user browser after it closes—does not establish the required control for a nonprofit grant portal to decouple customer event ingestion from slow downstream GenAI processing. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a nonprofit grant portal. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—block the transaction until every AI job finishes—does not establish the required control for a nonprofit grant portal to decouple customer event ingestion from slow downstream GenAI processing. The change is applied outside the component responsible for the requirement identified by a nonprofit grant portal. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 8
At a multilingual contact center, engineers need to provide a lightweight stateless capability to a remote agent. The service must continue to support multiple independent tenants. The initial assessment isolates this issue from unrelated infrastructure changes. Which change should the team prioritize?
- Expose an authenticated MCP-compatible tool endpoint backed by Lambda where appropriate
- Grant the agent a blanket administrative session
- Build a separate foundation model for each tool
- Store persistent state only in a Lambda process global
Explanation
Correct Answer
A:MCP describes interoperable tool access; runtime deployment still requires secure authentication and suitable state handling. In a multilingual contact center, the engineering team should verify this design against its stated objective to provide a lightweight stateless capability to a remote agent, including the requirement that the service must continue to support multiple independent tenants.
Incorrect Answers
B:The suggested action—grant the agent a blanket administrative session—does not establish the required control for a multilingual contact center to provide a lightweight stateless capability to a remote agent. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a multilingual contact center. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—build a separate foundation model for each tool—does not establish the required control for a multilingual contact center to provide a lightweight stateless capability to a remote agent. It replaces a verifiable engineering control with an untested shortcut in the a multilingual contact center deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—store persistent state only in a Lambda process global—does not establish the required control for a multilingual contact center to provide a lightweight stateless capability to a remote agent. The change is applied outside the component responsible for the requirement identified by a multilingual contact center. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 9
At a wholesale distribution platform, engineers need to keep an entire policy section together across chunk boundaries. The architecture must minimize avoidable data exposure. The team wants a solution that can be validated through targeted test cases. What is the most appropriate production approach?
- Split every document into single-character chunks
- Use randomly sized chunks without overlap or structure
- Increase model output tokens without improving context selection
- Use structure-aware or hierarchical chunking and test answer recall
Explanation
Correct Answer
D:Structure-sensitive chunking preserves relationships between headings and text. In a wholesale distribution platform, the engineering team should verify this design against its stated objective to keep an entire policy section together across chunk boundaries, including the requirement that the architecture must minimize avoidable data exposure.
Incorrect Answers
A:The suggested action—split every document into single-character chunks—does not establish the required control for a wholesale distribution platform to keep an entire policy section together across chunk boundaries. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a wholesale distribution platform. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
B:The suggested action—use randomly sized chunks without overlap or structure—does not establish the required control for a wholesale distribution platform to keep an entire policy section together across chunk boundaries. It replaces a verifiable engineering control with an untested shortcut in the a wholesale distribution platform deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
C:The suggested action—increase model output tokens without improving context selection—does not establish the required control for a wholesale distribution platform to keep an entire policy section together across chunk boundaries. The change is applied outside the component responsible for the requirement identified by a wholesale distribution platform. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
Question 10
At a digital publishing business, engineers need to limit prompt and response retention for a regulated application. Response reliability is more important than free-form creativity. The team wants a solution that can be validated through targeted test cases. What should be implemented before scaling this workflow?
- Assume disabling application logs eliminates all copies in supporting systems
- Write sensitive prompts to browser analytics automatically
- Set data retention and logging policies consistent with the compliance mandate
- Keep verbose unredacted invocation logs forever
Explanation
Correct Answer
C:Privacy controls must address the full data lifecycle, including diagnostic stores. In a digital publishing business, the engineering team should verify this design against its stated objective to limit prompt and response retention for a regulated application, including the requirement that response reliability is more important than free-form creativity.
Incorrect Answers
A:The suggested action—assume disabling application logs eliminates all copies in supporting systems—does not establish the required control for a digital publishing business to limit prompt and response retention for a regulated application. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a digital publishing business. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
B:The suggested action—write sensitive prompts to browser analytics automatically—does not establish the required control for a digital publishing business to limit prompt and response retention for a regulated application. It replaces a verifiable engineering control with an untested shortcut in the a digital publishing business deployment. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.
D:The suggested action—keep verbose unredacted invocation logs forever—does not establish the required control for a digital publishing business to limit prompt and response retention for a regulated application. No demonstrated capability in this scenario guarantees that this step resolves the concern raised by a digital publishing business. In this deployment, engineers should verify the missing mechanism rather than assume the proposed step satisfies the acceptance conditions.