Rocket Money scales transformer model via Hugging Face Inference API for 4000+ classes in production

AI Impact Summary

Rocket Money moved from regex-based transaction normalization to a BERT-family classifier to support 4000+ classes in production, outsourcing hosting and scale concerns to Hugging Face Inference API. This enabled a rapid ramp and higher confidence in classification under bursty loads, leveraging GCP for training and Vertex Pipelines for model workflows. The rollout exposed reliability chokepoints—outages during class expansion and caching issues during model handoffs—highlighting the need for end-to-end telemetry, data freshness guarantees, and robust monitoring to maintain latency and downstream enrichment. The business impact includes better retention and engagement but rising inference costs as monthly transaction volume grows, driving a cost-performance trade-off that will influence future scaling decisions.