Evaluation-First AI Agents: How Zepto Scales Customer Support on Databricks and MLflow
Databricks, Wednesday, September 9th, 2026
Zepto built evaluation before deployment, using MLflow to keep support agents reliable as they scaled.
Zepto, one of India's fastest-growing commerce platforms, built its customer support agents evaluation-first, establishing measurement before scaling rather than after problems appeared.
The case study covers using MLflow to track agent performance and catch regressions as prompts, models and tooling changed.
The evaluation-first sequencing is the transferable lesson, since most organizations deploy agents and then discover they have no way to tell whether a change improved or degraded behavior, at which point building evaluation requires reconstructing ground truth retroactively.