The gap between building AI and running it. Production notes on ML, data science, and the systems in between — handoffs, evals, cost, and what actually survives contact with users.
Rethinking Bronze, Silver, and Gold for the AI era
Press enter or click to view image in full size
Medallion Architecture Reconsidered
The medallion architecture — Bronze, Silver, and Gold — has become the default mental model for organizing a lakehouse, and it earned that position. But the critique of the pattern has grown sharper, and AI workloads have fundamentally changed what we ask of data platforms. In Part 1, we revisit why medallion emerged, identify exactly where it is cracking, and propose the redefinition these cracks demand.
Press enter or click to view image in full size
A Quick Refresher: What Medallion Architecture Is
Databricks documentation defines the medallion architecture as a data design pattern that organizes data logically, with the goal of incrementally improving structure and quality as data flows through each layer:
Bronze (Raw): Landing zone for source data as-is, with ingestion metadata. No business logic — its job is completeness and replayability.
Silver (Validated): Cleaning, deduplication, and conformance. The shared, trustworthy foundation for analysis.
The gap between building AI and running it. Production notes on ML, data science, and the systems in between — handoffs, evals, cost, and what actually survives contact with users.
In every article that describes how something should be done properly and how to change the current situation, I find a fundamental section is missing: HOW. How to move from one process or system to another. What to do first, what to do next. They…
Multi layer decision spport architectures have been around forever. Databricks just co opted it. Unstructured docs? Just put them into bronze... or anywhere really.
Vector DB's? Haven't we already evovled past needing to care about these, they are…
