oduonye/supportflow-banking77

SupportFlow Banking Intent Router

Intended use

SupportFlow predicts the operational intent of an English customer-support message so a queue, help-desk workflow, or human agent can route it faster.

It is not a bank, does not move money, and must not make credit, fraud, identity, or eligibility decisions. Low-confidence predictions should go to manual review.

Training data

• Dataset: PolyAI/banking77
• Source: PolyAI task-specific datasets
• License: CC BY 4.0
• Task: 77-way English banking-support intent classification

Base model and training

The pipeline fine-tunes distilbert/distilbert-base-uncased with a deterministic seed, a stratified validation split, and the untouched upstream test split for final evaluation.

Limitations

• The examples are English and banking-specific.
• The model can be wrong when a message contains several issues or a new intent.
• Dataset labels and customer language may not match a particular company's support taxonomy.
• Accuracy and F1 are benchmark measures, not a guarantee of safe production behavior.