A privacy-first local-LLM pipeline for weekly support trends

Problem

The helpdesk offered no built-in way to monitor topic or trend patterns at our plan tier, so there was no systematic view of what merchants kept asking about or which feature requests were gaining momentum.

Finding

Support conversations could be sorted into useful review buckets for feature requests, integration questions, churn signals, and subscription changes, giving a weekly read on what was actually trending.

Resolution

The pipeline (Python + JavaScript, with glm-4 running locally via Ollama over the de-identified per-conversation text files) classifies the prior week’s conversations and presents the team an at-a-glance summary. It’s a heads-up instrument, not an autopilot — the team decides whether any trend warrants action.

Impact

One consistently surfaced request was conversion tracking for non-Stripe payments; in response, the team documented and openly exposed the non-Stripe conversion-tracking API endpoints for merchants with dev teams. The whole flow — extract, de-identify, classify — keeps customer data on my machine end to end.