ArthSetu is a multi-agent AI lending orchestration system that automates the complete personal loan journey, from customer conversation, KYC verification, underwriting, and sanction letter generation to cross-selling and fraud detection.
Unlike traditional chatbots, ArthSetu uses a Master Agent + multiple Worker Agents to autonomously handle lending decisions using deterministic, transparent and RBI-aligned business logic.
ArthSetu bridges people and finance using explainable, tokenized, and RBI-ready conversational intelligence & not opaque black-box credit scoring.
- Master Agent controls state and routing
- Worker Agents execute:
- Sales negotiation
- KYC verification (CRM/eKYC)
- Underwriting rules
- Sanction letter generation
- Cross-selling if rejected
- Anomaly and fraud risk flags
- All loan amounts, EMI, interest and approvals are pure code logic, not LLM predictions
- GenAI controls tone & persuasion, not numeric decisions
- Aadhaar, phone, KYC docs, and salary slips are encrypted and tokenized
- Agents can only access tokens, never real data
- Auto-purge after loan lifecycle = DPDP-Act compliant
- Approval decisions are deterministic:
- Credit score thresholds
- Pre-approved limit matching
- Salary-to-EMI ratio
- Every decision is transparent and auditable
- Built for multilingual onboarding: Hindi, Marathi, Tamil, Bengali, Telugu
- Smart mid-conversation language switching
- Fully voice-enabled for semi-literate and visually impaired users
- React / Next.js
- Whisper — Speech to text
- MarianMT — Text translation (Indian languages ↔ English)
- GTTS — Text to speech audio generation
- FastAPI — REST orchestration
- LangGraph — Multi-agent workflow engine
- Tesseract — Salary slip OCR
- Redis — Session state for conversational memory
- PostgreSQL — Main DB + encrypted vault storage
- Fernet — Column-level encryption for PII
- ChromaDB — Context retrieval for user help/explanations
- CRM lookup
- OfferMart loan catalogue
- Credit bureau score lookup
- Aadhaar/eKYC simulation
| Agent | Responsibility |
|---|---|
| Sales Agent | Negotiates loan terms, generates best offers, adapts tone |
| Verification Agent | CRM lookup, fallback to Aadhaar eKYC, fraud flags |
| Underwriting Agent | Score-based approval, DTI check, loan eligibility |
| Sanction Agent | Official sanction letter generation (English PDF) |
| Cross-Sell Agent | Alternate instant loan recommendations |
| Anomaly Agent | Behavior risk detection, re-authentication triggers |
- Raw PII → stored encrypted in vault
- Only safe tokens are exposed to agents
- Minimizes data blast radius
- Data automatically removed after completion
- Aligns with DPDP Act (India)
- Every state transition and decision is logged with integrity
- Lending trail is immutable and verifiable
- RBI digital lending compliant
- If anomalous behavior is detected:
- session lockdown
- risk scoring
- OTP re-authentication
- Improves fraud resilience
- Rules instead of opaque ML risk models
- Eliminates bias by design (gender, region, language)
- Meets RBI/SEBI ESG lending mandates
| Strategic Moat | Differentiator |
|---|---|
| Voice-First Vernacular UX | Enables lending access for 300M+ semi-literate users |
| Agentic AI Orchestration | Autonomous workflow = faster processing, higher approvals |
| Tokenized Zero-Trust Vault | Agents never see PII → huge compliance moat |
| Explainable Rule Engine | Transparent decisions for underwriting + audit readiness |
| RBI/DPDP Alignment | Lower regulatory risk and easier certification |
| Cross-Sell Intelligence | No dead-end journeys → improved conversion & retention |
- Faster processing: reduced manual dependency
- Higher loan completion rates in vernacular markets
- Seamless self-serve onboarding replacing physical counters
- Lower fraud exposure and compliance penalties
- Full traceability for credit decisions
Backend designed to run with:
- FastAPI app server
- Redis for session state
- PostgreSQL for encrypted vault and audit logging
- React/Next.js frontend for user interface
Every borrower can speak their way into financial inclusion.