At AI Engineer NYC on October 13, Bridgewater’s Aaron Linsky and Sam Green described Vault, its unstructured-data platform. The team hand-labeled about 200 search questions, including misleading near-matches, in under a week. Linsky reported precision among the top ten results rising from about 60% to 90–95% after improvements culminating in agentic search. Adoption also required translating familiar email subscriptions and folders into saved searches and collections. Worth watching for the implementation details: talk and transcript, 16:21. Livestream markers: 18:47 comparison, 23:13 evaluation, 25:08 filtering. Assessment is from the publisher’s reading edition and transcript, not listening.