Zepto Scales AI Customer Support with Evaluation-First Agents on Databricks & MLflow

This article details Zepto's methodology for building reliable and scalable AI customer support agents, which manage over 80% of their daily tickets. It showcases an "evaluation-first" approach leveraging Databricks and MLflow for robust testing, tracing, and LLM-as-judge evaluations. The framework, based on a dual-loop architecture, has led to a 65% reduction in support costs, improved customer satisfaction, and faster development cycles. Zepto's experience provides a reusable blueprint for engineering high-impact, production-grade AI agents.

• →Fixes were slow
• →Build golden datasets as an asset, with owned KPIs (size, dev–prod gap, edge-case coverage)