Voice AI for Outbound Collections

Overview

I led the development of a production Voice AI system for outbound collections at a large multi-country commerce platform. The system is currently operating across three countries, has improved recovery performance over the previous baseline, and delivers positive ROI.

The challenge

Every call had to satisfy strict business rules: confirm eligibility immediately before dialing, follow a governed conversation policy, produce structured outcomes, and write reliable actions back to a legacy CRM. That CRM was the hardest integration point because its APIs were inconsistent and had limited resilience.

From research to production

I began with Pipecat and LiveKit prototypes that business teams could test directly. Once the opportunity was validated, I coordinated requirement discovery, delivered an MVP, and tested it with real users. Those results justified the design and implementation of a scalable production architecture.

The production system

Genesys manages campaign orchestration and dialing, while ElevenLabs powers the real-time voice agent. We also evaluated an internal Pipecat-based alternative during the research phase.

After each call, a secure gateway validates and queues the event. The processing layer normalizes the outcome, evaluates the conversation, extracts structured data, and records activities and validated payment promises in the CRM.

Reliability and observability

I designed the integration around asynchronous processing, durable queues, schema validation, timeouts, controlled retries, deduplication, idempotency, reconciliation, and operational alerts. These controls prevent business actions from being lost or duplicated when dependent systems partially fail.

My role

As lead developer and technical owner, I worked directly with the business team to understand the operation and translate it into product and technical requirements. I built the conversational agent, evaluation and observability flows, internal post-call processing, CRM integration, and Genesys integration.

Results

• Operating across three countries • Recovery metrics above the previous baseline • Positive ROI

This project shows how I approach Voice AI for real operations: validate the user and business value early, then build the surrounding integrations and controls with the same care as the conversational experience.