Can Your AI Analyst Keep a Secret? - Aaron Linsky, Bridgewater & Nikhil Benesch, turbopuffer

Why watch

This talk explains how Bridgewater Associates solved searching across 100 million permissioned documents using turbopuffer's native filtering approach. The solution addresses the tradeoff between post-filtering (loses recall) and pre-filtering (too slow) by implementing filtering directly in the vector index, achieving ~120ms P90 latency across 400 million vectors. Ideal for engineers building AI applications with permissioned data or anyone scaling vector search systems.