Scaling dbt on Amazon Redshift: how KOHO cut transformation runtime 70% without rewriting a single model
KOHO's data leaders argue that self-serve analytics with AI only works once you've already built the data culture and the governed dbt foundation underneath it.
Overview
KOHO Financial, serving 2.5 million customers, scaled to 200 TB of data and 30 million daily events, pushing nightly dbt jobs past 6.5 hours. Covers redesigning from centralized to Hub & Spoke to a data mesh on Redshift Serverless, cutting runtime ~70% (6.5h to 2h) through configuration alone with no model rewrites.
What you'll take away
- 01
The hard part of self-serve analytics was never the AI: years of data-team culture work and foundational governance made KoDa, KOHO's Claude-powered tool, possible at all.
- 02
KoDa is deliberately scoped: analytics use cases only, certified dbt mart models only, no PII, read-only, and every answer traces back to the same company-wide dbt definitions.
- 03
The real engineering effort went into evaluation, not the model: an "eval council" built a library of about 375 real analyst questions to catch regressions before they reach users.
Session speakers
Manikandan Paramasivan
KOHO FinancialRaza Hafeez
Amazon