Focus
Datasets
Corpora, label schemas, and the adversarial cases that decide whether a model is honest.
RAG quality is a data problem first. So is compliance screening, drift detection, and anything that claims to know the law or a tender notice. The schema you choose decides what the model is even able to be honest about.
I don't have a dataset published here right now. What's on this shelf instead: an agent that evaluates other people's datasets before you train on them, and the Prospecto pipeline, which is the closest thing I have to dataset engineering running in production.
On this shelf
Evidence-led dataset discoveryWork · live SpaceAn agent that reads dataset cards, schemas, splits and sample rows, then ranks on evidence rather than download count.OpenRAG quality is a data problem firstNotesChunking, labeling, and corpus hygiene before you touch the generator.OpenProcurement intelligence at national scaleWork · ProspectoDiscovery and extraction over 20,000+ government sources — unstructured notices into structured records.Open
EU AI Act & GDPR instruction setNextA compliance instruction corpus built on the rule packs behind Compliance Agent. Not started.Not built