Ryan Ahern introduces Exa Snapshot as a way to retrieve a known page or execute a search against a past point in time. He connects this to agent evaluation: searching today's web can reveal an acquisition or other outcome that a historical prediction was supposed to anticipate. Restricting retrieved evidence helps separate genuine predictive signals from information that arrived later.
Ryan Ahern describes three patterns: enriching an existing model with historical unstructured data, adding an agentic research feature to a conventional model, and testing a fully agentic prediction system. Financial examples illustrate the evaluation problem rather than establish a profitable trading method. Product coverage and performance claims are attributed to the presenter, not independently validated here.
Ryan Ahern uses an earlier workshop's company-name shortcut to show that historical retrieval alone does not prevent overfitting. Teams still need held-out examples and checks for generalization. The short talk is distinct from the separate hands-on workshop: it organizes the use cases and explains why the information available at prediction time matters.
Watch on YouTube
