$1B in transactions, 1 minute latency: dbt as production banking infrastructure

Keep is trusted by 5,000+ Canadian businesses processing $1B+ in card volume, and dbt sits underneath both its analytics and its live credit decisions. When a customer applies for credit, banking data streams into Snowflake, dbt models refresh on a one-minute lag, and an AI model scores the application and generates a credit recommendation for compliance review within minutes. With fresh data instead of stale batches, our system catches risk signals a slower pipeline would miss entirely, allowing us to confidently extend credit faster.

In this session, we'll cover the architecture behind Keep’s production AI system, including:

• Low-latency modeling patterns
• Safe model design for production workloads
• Cost, freshness, and reliability tradeoffs

We'll also show how dbt feeds financial data into accounting systems to keep a live view of the book of risk, and how snapshots and slowly changing dimensions let us time travel through those tables for audits and compliance.

You'll leave with a pattern for building dbt-powered systems fast enough for AI decisions, and durable enough to hold up under scrutiny.