Vinoo Ganesh demoed Kepler verification after $270M flowed to rivals, per Ganesh. The demo covered RAG platforms and agent pilots that read SEC filings and internal notes. Attack needs only remote access. A missing tenant filter or a spoofed caller identity exposes another firm, per AI Engineer.
Three months after the AIE World Fair talk framed verification as a required trust layer, Ganesh returned with architecture, per AI Engineer. Anthropic had published an explainer on how Kepler works, per Ganesh. The shift is from post hoc checks to preprocessing where code writes pointers to facts. The earlier warning focused on models writing numbers, now models cannot print without dereferencing.
Ganesh, Kepler CEO and former Citadel business engineering head, rejects the idea models will outgrow error, per AI Engineer. He cites 80% of finance execs trusting numbers while 31% find errors, per Ganesh. His line was "Alpha is how a firm makes money." That turns creativity into edge and leaves generic RAG missing firm data.
Lookup cannot miss like search can, per Ganesh.
Kepler pulls up to 200 candidates with ANN plus BM25 then narrows to 60 with reciprocal ranking and cross-encoders, per Ganesh. Code pulls facts by pointer dereference and refuses when a pointer fails, with refusals falling but not to zero since inference stays probabilistic, per Ganesh. Tenant scoping, allow lists and populated indexes block stale or cross firm reads, per Ganesh. A separate help centre for a product also named Kepler describes an app that turns what you find online into context for AI, per its docs. The next proof is a cradle to grave DCF grounded in 10K filings.
