Challenge
Consumer lending collections across multilingual cohorts suffered high delinquency roll-rates when outreach used manual dialers and English-only notifications, wasting agent capacity on self-curing accounts.
Technical Implementation
- DPD Propensity Segmentation: Modeled borrower transition probabilities across Days Past Due (
DPD) buckets to calibrate outreach timing and restructuring eligibility. - 28-Language MuRIL Intent Routing: Applied Google’s
MuRILtransformer encoder to classify inbound borrower replies—including code-mixed regional scripts—across 28 languages. - Serverless Omnichannel Dispatch: Automated two-way
WhatsApp BusinessandTwilioworkflows onAWS Lambdapaired with hierarchicalDjangosupervisor dashboards.
Production Outcome
Increased borrower repayment recovery rates by 37% while automating routine follow-ups and pre-approved credit cross-selling.