Intelligent customer onboarding portal and automated qualification system engineered for State Bank of India.
Overview
Sarathi AI is a conversational onboarding concierge that guides banking customers through product selection and KYC, dynamically adjusting the flow based on user responses and compliance rules.
The Problem
Traditional banking onboarding systems are slow and complex, often lacking conversational support, leading to high drop-off rates and insecure KYC validation.
System Architecture
Vite + React frontend with a Node.js orchestration backend. The backend manages a multi-agent system (Product Agent, KYC Agent, Compliance Agent) using LangChain.
Engineering Trade-offs
Chosen SSE over WebSockets for streaming to simplify firewall traversal in enterprise banking environments, despite a slight increase in connection setup overhead.
Lessons Learned & Future
Production Learnings
Multi-agent systems can easily get stuck in loops if they disagree. Implemented a strict 'Orchestrator' pattern to forcefully resolve deadlocks.
Future Improvements
Integration with real Aadhaar UIDAI staging APIs for authentic e-KYC validation testing.
Database Layer
MongoDB for flexible conversation state storage, allowing agents to append arbitrary context to a user's session.
API Design
Server-Sent Events (SSE) used to stream LLM responses to the frontend while simultaneously running compliance checks in parallel.
Security Decisions
1. The Compliance Agent acts as a strict output parser; if it detects PII exposure or regulatory violations, it overrides the conversational response with a hardcoded safe message.
Scale & Metrics
Maintains context across 20+ turn conversations with <1.5s time-to-first-token.