Customer support teams spend significant time reading and routing incoming requests.

I developed TicketTriage AI to explore how natural language processing can assist this workflow.

What the system does

  • Classifies support tickets by category and urgency.
  • Uses TF-IDF and Logistic Regression.
  • Returns confidence scores and token-level explanations.
  • Provides a FastAPI prediction service.
  • Includes a Streamlit interface.
  • Flags low-confidence predictions for human review.

Engineering workflow

The repository documents synthetic-data generation, model training, evaluation, automated testing, Docker configuration, and GitHub Actions.

Macro F1 is used as a primary evaluation metric because performance across different ticket categories matters.

Limitations

The training dataset is synthetic. Real support conversations contain more variation, ambiguity, and changing language patterns.

The system is intended to assist support agents, not replace human decisions.

GitHub repository: https://github.com/tasnem-tech/ticket-triage-ai

I'm particularly interested in improving model confidence and practical human-in-the-loop workflows.