Cluster 90k open-text survey responses into a product feedback taxonomy

Three years of NPS follow-up comments, in-app feedback, and churn survey answers are sitting in a spreadsheet nobody reads - about 90k free-text entries in English and Spanish. We build field service software and we know the answers to our roadmap debates are buried in there. What we need: an embedding-and-clustering pipeline that groups the responses into a stable theme taxonomy (we expect 30-60 themes), LLM-generated labels and representative quotes per theme, and a repeatable Python script our team can rerun quarterly as new feedback arrives. The final readout is a written report ranking themes by volume and by correlation with churned accounts. Short, focused project. We will judge proposals on whether you describe how you would validate cluster quality - if the plan is just "run k-means and eyeball it", it is not for us.