What is medallion architecture?
Medallion architecture is a convention for organizing the tables in a data lakehouse into three layers of increasing quality, named bronze, silver, and gold. Bronze holds data exactly as it arrived from the source system, including duplicates and errors, with a timestamp and a record of where it came from. Silver holds the same data after cleansing, de-duplication, type correction, and mapping onto shared definitions, so a customer identifier means the same thing regardless of the system that produced it. Gold holds tables shaped for a specific business purpose, such as monthly revenue by territory or a feature table for a churn model.
Each layer is built only from the layer below, by code that can be re-run. If a cleansing rule changes, silver and gold are rebuilt from bronze without going back to the source systems.
Why medallion architecture matters
The pattern separates three jobs that otherwise tangle: landing data reliably, making it correct, and making it useful. Engineers own bronze and silver, while analysts and domain experts shape gold. Data lineage follows naturally, because every gold figure can be traced to the silver rows and the bronze file that produced it. When a number on a board report looks wrong, the team walks down the layers instead of guessing.
AI projects benefit directly: models train on silver, where data is clean but not yet aggregated away, and agents and dashboards read gold, where definitions are stable and access is controlled.
Medallion architecture in practice
A collections agency receives placement files from twenty creditors in twenty formats. Each file lands in bronze untouched. A silver job parses every format into one debtor and account schema, standardizes phone numbers and addresses, and flags accounts that already exist. Gold tables summarize portfolio performance by creditor and feed the dialer’s prioritization model. When one creditor changed its file layout, only that creditor’s silver parser needed a fix, and gold was rebuilt within the hour.
In Kirk Tech Solutions engagements, the canonical schema is usually enforced at the silver layer, so every downstream model and agent reads one definition of each entity.
What to watch for
- Keep bronze immutable, because the ability to rebuild everything from it is the main safety net.
- Keep silver free of purpose-specific business logic so it stays reusable, and put those rules in gold.
- Define data quality checks at the silver boundary so bad records are quarantined before they reach reports.
- Expect the number of gold tables to grow quickly, and give each one an owner.