Voice of Customer: Unsupervised Insight Engine

What I built

Product managers spent 40+ hours per quarter manually reading and tagging support tickets to identify product issues. Unstructured feedback created a data bottleneck where systemic problems stayed hidden until customer churn revealed them. No quantitative way to prioritize which issues impacted the most customers.

Build an automated pipeline that converts qualitative customer feedback into quantitative strategic insights without requiring manual tagging or supervision, while ensuring PII compliance and cost predictability for enterprise deployment.

What I did

Architected a complete unsupervised learning pipeline: automated ETL ingestion with PII redaction (regex patterns), semantic embeddings via OpenAI, K-Means clustering to surface latent patterns, and LLM-powered cluster summarization. Implemented FinOps middleware to track and cap API costs per pipeline run. Built interactive dashboard with drill-down capabilities for cluster exploration and pattern analysis.

What came out of it

Processed 1000 tickets in under 2 minutes, reducing analysis time by 90%. Automatically discovered 15 distinct issue categories with human-readable summaries. Cost per run: $0.16 (embeddings + summaries), making it economically viable at scale. Dashboard enables product teams to instantly identify top issues, track trends over time, and prioritize roadmap items based on actual customer impact volume.