Rocket Money scales ML with Hugging Face

Rocket Money transitioned from its existing regular expression and decision table approach to a BERT-based model for classifying financial transaction history.

The main challenges were the sharp traffic fluctuations that occur while processing over 100 million transactions per month and the complexity of managing more than 4,000 classes. Without a dedicated MLOps team, instead of building infrastructure, they adopted Hugging Face's Inference API to solve the model serving problem.

This reduced the burden of infrastructure management and built an environment where ML engineers could focus on model performance optimization, while also overcoming compatibility issues with the GCP environment and high availability concerns.