Voice of the Customer — Feedback Pipeline for a Swiss Wholesaler
A Swiss wholesaler collects customer feedback across several touchpoints but evaluated it largely by hand. We built the pipeline that runs from raw feedback to a classified metric: export from the survey system, AI-assisted classification, database, reporting views, dashboard. The engagement is ordered in stages, and the code transfers to the client at the end.
Starting point
Feedback from various channels arrived as export files and was read, categorised and copied into reports by hand. That cost time, was not repeatable and produced no time series: two people categorised the same feedback differently, and nobody could evidence which reading was right.
What we built
An end-to-end processing pipeline: read the export, classify the feedback with a language model, store it in a PostgreSQL database, layer reporting views on top and surface a dashboard for the business units. Plus an operations console for starting runs, inspecting errors and re-running individual steps. Everything containerised and operated on a Swiss server.
Quality measured, not asserted
Each metric has its own gold set: feedback classified by hand by domain experts, against which the machine classification can be measured. That is the difference between «the AI categorises» and «we know how well the AI categorises» — and it is the precondition for swapping the language model later without silently changing the analysis.
Ordered in stages, built for handover
The engagement was ordered and delivered in stages. Stage 0 settled feasibility and acceptance criteria, stage 1 delivered the end-to-end slice — the path from raw file to classified feedback ran through completely in production for the first time in August 2026. Shipped from the outset: an operations manual with start commands, exit codes and logging, plus a clear split of what transfers to the client.