Business Context

ZendeskPro, a B2B support platform, wants to modernize its ticket-routing pipeline. The team currently uses keyword rules, but they want an NLP solution that compares common algorithm families used in practice and recommends the best approach for production.

Data

You are given 180,000 historical support tickets collected over 18 months. Each ticket contains a short subject line and a free-text description. Text is primarily English (96%), with small amounts of Spanish and French. Ticket length ranges from 8 to 420 words (median: 54). Labels represent 6 routing queues: Billing, Technical Issue, Account Access, Feature Request, Bug Report, and General Inquiry. The class distribution is moderately imbalanced, with General Inquiry making up 34% of the data and Account Access only 7%.

Success Criteria

A good solution should clearly compare common NLP algorithms across classical and modern approaches, achieve macro-F1 >= 0.82, and keep p95 inference latency under 120 ms per ticket in batch serving.

Constraints

• Must run in a secure Python environment using open-source libraries only
• Training budget is limited to a single T4 GPU or CPU fallback
• The final recommendation should balance accuracy, interpretability, and deployment simplicity

Requirements

1. Build a baseline using TF-IDF + Logistic Regression.
2. Build a second baseline using averaged word embeddings + linear classifier.
3. Fine-tune a lightweight transformer for the same task.
4. Define a realistic preprocessing pipeline for noisy support text.
5. Compare the algorithms on accuracy, macro-F1, latency, and operational trade-offs.
6. Explain when each algorithm family is appropriate in production NLP systems.