AI / NLP / Internship project
Different languages. Shared context.
A multi-task pipeline that understands, classifies and prioritizes support tickets in three languages.
The problem
Imbalanced classes and language confounding in synthetic Turkish data made reliable evaluation difficult.
The approach
I compared TF-IDF baselines with BERT and XLM-RoBERTa, using customer-grouped splits, focal loss and consistent labels. I served the model through FastAPI microservices.
Results and scope
- 32.7K support tickets · EN / DE / TR
- Project evaluation: 0.85 type macro-F1; 0.63 mean macro-F1 across four tasks.
- Separate internship XLM-R experiments: approximately 0.87 category and 0.78 priority macro-F1. These scores belong to different experiments.