Customer Challenge 

A leading consumer lending company serving millions of borrowers was running its customer support operation almost entirely with human agents — driving high operational costs and long wait times for repetitive queries on loan eligibility, application status, EMI schedules, and document requirements. Existing chat solutions lacked intelligence, personalisation, and voice channel support. The customer required an agentic AI platform that could autonomously understand intent in Hindi, English, and Tamil, retrieve contextually relevant financial product information, personalise responses using live customer profile and loan data, execute complex eligibility business rules, orchestrate document collection, and escalate to human agents — all within strict RBI and DPDPA compliance and a sub-4-second response latency SLA across chat and voice. 

Solution 

Aivar designed and deployed a fully serverless, multi-agentic conversational AI platform on AWS built around an orchestrating AI agent powered by Amazon Bedrock (Claude Sonnet 3 Haiku and Claude Sonnet 3.7) running on AWS Lambda and Amazon ECS Fargate. The orchestrator coordinates five specialised AWS Strands tools — a Customer Data and Profile Agent backed by Amazon DynamoDB with PII redaction at ingress, a Business Rules Engine for eligibility and compliance validation, a Response Generator grounded in a FAISS RAG knowledge base on Amazon S3 with Amazon Titan Embeddings V2, a Document Requirements and Collection Agent, and a Support Escalation Agent with full context handoff. Voice channels are served by a real-time pipeline combining Deepgram STT (fine-tuned for financial terminology), the same agent toolchain, Sarvam TTS for Hindi, English, and Tamil, and Livekit for delivery — all within the 4-second end-to-end latency SLA. The entire environment is provisioned with Terraform and deployed via AWS CodePipeline and CodeBuild with Checkov security scanning; security and compliance are enforced through least-privilege IAM, KMS encryption across S3 and DynamoDB, AWS Cognito MFA, CloudTrail audit logging, and VPC Flow Logs. 

Architecture 

The architecture combines AWS serverless, AI, data, and security services for enterprise-grade availability and compliance — including AWS Lambda and Amazon ECS Fargate for orchestration and tool execution, Amazon Bedrock (Claude Sonnet 3 Haiku and 3.7) for the agentic reasoning layer, AWS Strands for tool coordination, Amazon DynamoDB for unified channel-agnostic session state, Amazon S3 with a FAISS index and Amazon Titan Embeddings V2 for the RAG knowledge base, Deepgram STT, Sarvam TTS, and Livekit for the real-time voice pipeline, Amazon Comprehend for PII redaction, AWS Cognito for authentication and MFA, AWS WAF for edge protection, AWS KMS for encryption, CloudTrail for audit logging mandated by RBI, OpenTelemetry distributed tracing across all tool invocations, and Amazon CloudWatch with Amazon QuickSight dashboards for deflection, escalation, cross-selling, and multi-language routing observability. 

Key Outcomes 

• 65%+ bot deflection — majority of standard loan lifecycle queries resolved autonomously across chat and voice without human agent involvement 

• Sub-4-second voice SLA — end-to-end voice pipeline (STT → agent → TTS → delivery) executed within the 4-second latency target 

• Multi-language autonomy — accurate language detection, routing, and voice synthesis across Hindi, English, and Tamil 

• 100% policy compliance — Business Rules Engine ensured policy-compliant responses for financial product eligibility queries with no human review for standard cases 

• Document collection automated — high-volume document requirement and collection workflows handled autonomously, eliminating the largest category of manual support interactions 

• Full RBI and DPDPA compliance — 100% of customer conversations processed through PII redaction before storage, with CloudTrail covering all conversation storage and retrieval operations 

• Cross-sell uplift — contextual nudge engine autonomously surfaced relevant product offers at post-repayment and post-disbursement milestones, contributing to incremental revenue 

• Seamless escalation handoff — Support Escalation Agent passed full conversation context to human agents, eliminating customer repeat-explanation