2025–2026
Tell me about the AI collections work.
AI-assisted collections over WhatsApp for banks
The problem
Banks wanted to handle overdue payments over WhatsApp without a person on every conversation: remind the customer, agree a promise to pay, receive proof of payment and measure satisfaction, while staying safe when the AI isn't sure.
What I built
- Created the preventive-collections bot: opening message plus WhatsApp Flow, promise-to-pay flow, an "already paid" path with receipt upload, date and amount validations and a star-rating survey.
- Integrated an LLM collections agent and a knowledge-base FAQ with function calling. When the model's confidence is low, the bot shows a disambiguation menu instead of guessing.
- Message-burst buffering: groups messages a customer sends in quick succession before the AI answers.
- Receipt OCR inside WhatsApp Flows with OpenAI vision for Invex and a multi-client payment-report flow: detects the issuing bank, rejects handwritten receipts and enforces a hard timeout.
- Banorte: collections-bot features (CSAT, deeplinks, AI-agent handling), the CSAT survey's persistence layer with Meta's new decryption method, and collections API endpoints.
- Banco Azteca service-rating survey: moved its cache from memory to Redis/ElastiCache.
My role
I created the preventive-collections bot and the receipt-OCR features, and contributed to the Banorte services alongside the finance team.