Eugene Yan’s latest article, “More Design Patterns For Machine Learning Systems,” adds nine patterns to the vocabulary of ML practitioners. The collection spans data quality, training, and decision logic: human-in-the-loop systems, hard negative mining, reframing, cascades, data flywheels, and a business rules layer, among others. Together, these patterns address problems that often emerge after a model enters production.
The article presents hard mining as a way to focus training on difficult examples, and data flywheels as a loop that uses production feedback to improve future datasets. The cascade pattern suggests starting with a cheaper model and escalating harder cases to a more powerful one, while human-in-the-loop makes human review an explicit part of the pipeline. The business rules layer, meanwhile, lets teams combine deterministic checks with model output rather than relying on prediction alone.
Because the source is a single essay, the patterns are not compared head-to-head or ranked; each is presented as a practical option for specific circumstances. The value is in having a shared vocabulary for common design choices in ML systems.