Customer Data engAnalyticsLLM / RAG retailtelecomservices Support deflection +19%

Voice-of-customer engine

3 people · 5 months Data engineer · ML engineer · Me — architecture and delivery
the engagement client data — not shown
heard
"Every team reads a different slice of the feedback and they all come to a different conclusion."
real problem

Free text arrived from six channels in four languages with no shared taxonomy. Volume was read as importance, and the tickets that actually drove cost were invisible.

system built

An ingestion layer that normalises and de-duplicates feedback, a retrieval-backed classifier onto a maintained issue tree, and a weekly digest that ranks issues by cost-to-serve rather than mention count. Every claim links back to source quotes.

what broke

The issue tree drifted. We introduced a monthly review where the top unclassified cluster either becomes a node or is explicitly rejected.

the open rebuild same architecture · public data
dataset

Amazon Reviews 2023 + Multilingual Amazon Reviews Corpus

what the rebuild covers

Classifier, retrieval index and the cost-weighted ranking are reproduced. The digest is a notebook, not a service.

artifacts
NotebookCodeDatasetWrite-upDemo
Open the code Open notebook