Automated enterprise data migration with AI reasoning for missing translations and data.

A client needed to extract more than 10,000 products and recipes from one CMS and load them into another. The records were bilingual but inconsistent: French descriptions missing for some products, English missing for others, unit conventions that didn't reconcile cleanly (imperial in one record, metric in the next, mixed within a single recipe), and category mappings that had drifted over years of edits.

The deterministic parts were straightforward: extract, transform fields with known shapes, load. The hard part was the cases where a clean rule couldn't decide. A French product description might be missing because no one wrote one, or it might be missing because the English description was used as a placeholder years ago. The right answer for the migration differed depending on which it was, and there were thousands of these.

I built the migration with an AI judgment layer specifically for those decisions: the deterministic pipeline ran first, and AI handled the residual cases with the context to decide reasonably. Output was reviewable before commit, with the AI's reasoning logged alongside each decision. The migration completed cleanly, and human review was confined to the records that needed a judgment call.