SecureBank Assist
(Enterprise Dialogflow CX Agent)
A production-grade Dialogflow CX virtual agent for a fintech contact center covering OTP auth, account services, and complex transactions.
The Challenge
Enterprise conversational AI requires handling complex state machines where users constantly change their minds, interrupt workflows, or provide invalid inputs. Linear chatbot scripts break immediately in production.
SecureBank Assist was built as a true state machine using Dialogflow CX. I implemented advanced conversational design patterns including three-stage fallback strategies, contextual reprompting, and a robust hub-and-spoke router (Main Menu) that allows users to seamlessly switch contexts globally without breaking the flow.
System Architecture
stateDiagram-v2
[*] --> StartFlow
StartFlow --> AuthFlow: Collect ID
AuthFlow --> OTPFlow: Verify OTP
OTPFlow --> MainMenuFlow: Success
MainMenuFlow --> AccountServices
MainMenuFlow --> Transactions
MainMenuFlow --> CardServices
MainMenuFlow --> LoanServices
AccountServices --> Webhook[Backend Webhook]
Transactions --> Webhook
state GlobalCommands {
[*] --> Cancel
[*] --> MainMenu
[*] --> AgentTransfer
}
Key Features
Hub-and-Spoke Routing
Users can execute global commands (repeat, main menu, human agent) from anywhere, ensuring a non-linear, flexible conversation state.
Three-Stage Fallback
Graceful failure handling via sys.no-match-1/2/default cascading from rephrasing, to examples, to human escalation.
PII Masking
Security first: Strict redaction rules applied to sensitive entities like account numbers, phone numbers, and OTPs.
Generative Fallback
Integrated Knowledge Base + Generative Fallback for unstructured queries, wired through agent.json.