Support Ticket Triage with NLP

Built a multi-label text classifier that reads each incoming ticket, tags intent and product area, and predicts urgency. Fine-tuned a compact transformer on 240K historically labelled tickets, distilled it for low-latency inference, and served it behind a FastAPI endpoint with confidence thresholds that fall back to human routing when the model is unsure. Added an active-learning loop so agent corrections are folded back into the next training run.