Transaction classification is a deceptively hard NLP problem. "AMZN*AB3F9D2 SEATTLE WA" needs to map to "Cloud Infrastructure" for an engineering team and "Office Supplies" for a procurement team. Context matters, and context changes per workspace.
The Pipeline
We run each transaction through three stages: vendor normalization (stripping payment processor noise), base classification using a fine-tuned model on 2M labeled financial transactions, and workspace-level fine-tuning based on historical corrections from that tenant.
The Feedback Loop Is Everything
Our base model starts at ~82% accuracy on cold accounts. But every time a user corrects a category, that correction feeds back into the workspace model. By the 60th correction, workspace-level accuracy typically exceeds 94%, and stays there.
“The best classification system is one that gets smarter the more you use it, not one that asks you to configure it up front.”
Ndifoin Hilary
Founder