Interview question
Review a managed AI architecture for region, identity, retention, and portability
Reviews a managed AI proposal against data flow, identity, region, retention, quota, observability, portability, and recovery requirements.
TL;DR
Reviews a managed AI proposal against data flow, identity, region, retention, quota, observability, portability, and recovery requirements.
Architecture review, managed responsibility boundaries, data-flow analysis, identity, residency, retention, service limits, and exit strategy.
Practice the problem like a real interview: restate, reason, implement, and test.
Review a proposal that sends EU customer documents through a managed model endpoint, managed vector store, hosted agent runtime, and third-party tool connector. The diagram lists product names but not data paths, identities, retention, regional behavior, quotas, or recovery. Produce the questions, required changes, decision record, and validation plan needed before approval.
Browser -> application API -> managed agent runtime -> managed retrieval -> model endpoint
Agent runtime -> third-party ticketing connector
Unknowns: where uploaded files and traces persist, which identity each hop uses, whether every feature is available in-region,
what happens at quota exhaustion, and how workflows move if the agent runtime is replaced.
architecture_decision:
data_classes:
customer_document:
allowed_regions: [eu]
model_training: prohibited
content_telemetry: prohibited
retention_days: 30
identities:
app_to_model: workload_identity:model-invoke
app_to_retrieval: workload_identity:tenant-scoped-search
agent_to_ticketing: delegated_connector:ticket-comment-only
failure_policy:
quota_exhausted: queue_or_degrade
regional_outage: no_cross_region_failover_without_policy_approval
portability:
domain_workflow_state: application_database
eval_dataset_and_results: exportable
provider_specific_features: [managed_grounding_guardrail]
validation:
- regional_endpoint_probe
- least_privilege_permission_test
- retention_and_deletion_test
- quota_and_outage_game_day
The review cost grows with dependencies, regions, data classes, and connector operations. A concise responsibility matrix prevents the same control from being assumed by everyone and owned by no one.
Next in AI Production Scenarios & Labs: AI production game day