03 / Secure Agentic ApplicationEnterprise delivery

BankingAgentic Helpdesk

My rolePrincipal GenAI + Backend Engineer

ContextFinancial services support

System classSecure Agentic Application

Portfolio evidenceArchitecture · decisions · production

01 / Brief

The problem

Support workflows needed faster access to accurate card and account information without turning sensitive customer context into an uncontrolled prompt. The system had to combine retrieval, relationships and actions while preserving identity and auditability.

02 / Constraints

Why it was hard

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.

03 / Architecture

System shape

The architecture is expressed as operational layers: experience and identity at the edge, bounded orchestration in the middle, governed data and tools underneath, and evidence across the entire path.

Banking / logical architectureSanitized portfolio view

01Secure APIFastAPI + JWT identity

02Intent layerCards / accounts routing

03Agent runtimeOpenAI Agents SDK + tools

04KnowledgePostgreSQL/pgvector + MongoDB + Neo4j

05GuardrailsPolicy, audit, fallback

04 / Decisions

The engineering judgment

Backend-owned orchestration

The agent runs inside authenticated application workflows rather than controlling identity or business state itself.

Hybrid context instead of one vector index

Vector retrieval supplies semantic evidence while graph relationships and structured data preserve account context.

Tools expose narrow business capabilities

Agents receive explicit, validated operations instead of direct database access.

Grounding and audit are part of completion

A response is not successful unless the system can explain its evidence and record the action path.

05 / Production

Hardening the system

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.

PythonFastAPIOpenAI Agents SDKRAGpgvectorPostgreSQLMongoDBNeo4jJWTGuardrailsTool callingHybrid retrievalBackend architecture

06 / Outcome

The delivered helpdesk reported a 40% reduction in mean ticket-resolution time while retaining secure, backend-controlled execution.

Next case study / 04