Your AI project is a data project wearing a hat

The model is a dependency. The pipeline is the thing you own.

Stalled AI work is almost never stalled on modelling. It is stalled because ingestion is manual, because nobody can reproduce last month’s index, or because there is no number that says whether quality moved.

Build the eval set before you tune anything. It does not have to be large — thirty examples you actually argued about beats three hundred you generated.

Then make the transform versioned and the ingestion observable. Once those two exist, model choice becomes a cheap experiment instead of a rewrite.