Contact-center NLP at Starbucks

Led the NLP track on the People Analytics team. The brief: turn unstructured contact-center text into something the HR and operations teams could act on without a human reading every ticket.

What I built

• Topic modeling with BERT embeddings + clustering, refreshed weekly. Surfaced the top emerging partner concerns before they hit volume thresholds in the existing structured reporting.
• Summarization with T5, fine-tuned on internal call summaries to match the in-house writing style. Cut analyst review time on long transcripts substantially.
• Semantic search over the historical ticket corpus, replacing keyword search that had been the standard tooling.

What I shipped vs. what I learned

The deployment was straightforward — the team had solid MLOps and Databricks infrastructure. The harder problem was getting downstream consumers to trust the output enough to act on it. Topic-model output is intrinsically fuzzy; people who are used to deterministic dashboards treat that fuzz as a defect, not a feature.

The pattern that worked: never show raw model output. Always pair every topic with three example tickets and a confidence band. The model was the engine; the explainability scaffolding was the product.

(Details abstracted for confidentiality. Available on request for relevant conversations.)