Classifies customer-service messages into 77 banking intents, then auto-routes
or escalates based on confidence. Runs entirely in your browser — the quantized
model downloads once and inference happens on your device.
92.2%test accuracy
98.9%accuracy when auto-routed
72%tickets auto-routed
90%fewer errors reaching customers
DistilBERT fine-tuned on 8,494 Banking77
examples, benchmarked against TF-IDF + logistic regression (88.7%) and a
from-scratch PyTorch classifier (88.0%) on a held-out test set of 3,076 messages.
Exported to ONNX and quantized to int8 — 4x smaller, no measurable accuracy loss.
Messages below 0.95 confidence are escalated; that threshold was
tuned on validation, never on test.