Principal AI systems engineering, platform delivery, and reliability
Secure Agentic ApplicationEnterprise delivery
BankingAgentic Helpdesk
My rolePrincipal GenAI and backend engineer
ContextFinancial services support
System classSecure Agentic Application
FocusIdentity, retrieval, guarded actions
Brief
The problem
Support workflows needed faster access to accurate card and account information. Sensitive customer context still had to remain bounded by identity, policy, and auditability.
Constraints
Security follows the request
01Card and account intents required different knowledge, tools and policy boundaries.
02Customer and support identity had to be carried from the API boundary into every agent action.
03Answers needed grounding across structured, vector and graph context.
04Tool calls required validation, bounded permissions, and failure handling.
Architecture
The security boundary
Identity and policy stay attached from the API boundary through retrieval and every tool call.
Banking / logical architectureSanitized view
01Secure APIFastAPI + JWT identity
02Intent layerCards / accounts routing
03Agent runtimeOpenAI Agents SDK + tools
04KnowledgePostgreSQL/pgvector + MongoDB + Neo4j
05GuardrailsPolicy, audit, fallback
Decisions
Decisions at the boundary
01
Backend-owned orchestration
The agent operates inside authenticated application workflows; identity and business state remain in the backend.
02
Hybrid context across retrieved and structured data
Vector retrieval supplies semantic evidence while graph relationships and structured data preserve account context.
03
Tools expose narrow business capabilities
Agents use explicit, validated operations with bounded capabilities.
04
Grounding and audit are part of completion
A response is complete only when the system can explain its evidence and record the action path.
Production
Controls around action
01JWT-backed request identity and authorization.
02Retrieval filtering for sensitive customer and product context.
03Timeouts, retries, and controlled fallback around external model/tool calls.
04Audit trails for tool invocation and final response.
05Evaluation of intent routing, groundedness, and schema-valid output.
Reported enterprise delivery outcome: a 40% reduction in mean ticket-resolution time, with identity and tool execution retained in the application backend.
Relevant problem?
Designing agents around identity, policy, and evidence?