MegaWahl — E-commerce data platform

Starting point

A growing Shopify shop stocking around twenty suppliers — each with its own catalogue format (CSV, XLSX, PDF), its own encoding, its own logic. Doing it by hand does not scale: creating products, translating, categorising, adding SEO — thousands of times, consistently, with no lost SKU numbers.

Brief

A data platform that automatically turns supplier catalogues into clean Shopify imports — project-based, as a standalone Python system alongside the shop.

Build

Eleven supplier processors (a shared BaseProcessor architecture, Polars) normalise the raw data. A transactional SKU manager assigns numbers with rollback (reserve → validate → commit/cancel) — no number is ever lost. An asynchronous translation pipeline (Polish→German) preserves HTML structure and protects non-translatables like SKUs and brand names. The core: an AI product categorisation across the live catalogue — embedding retrieval (BGE-M3) proposes candidates, a local vision model with constrained decoding decides (token masking: the model physically cannot invent a category outside the Google taxonomy), and every decision passes through a human review before it is written back to the shop.