Ryan Ahern demonstrates point-in-time retrieval and search with Exa Snapshot, then turns the idea into an acquisition-prediction exercise. Participants define yes-or-no features and test them against companies before an outcome was announced. The methodological aim is to keep future web information out of a historical evaluation, not to promise reliable financial predictions.
Ryan Ahern explains why an article's displayed publication date is not always a reliable historical boundary: pages can be revised. He distinguishes gathering historical features for an existing model from testing a research agent or an entirely agentic workflow. A time-bounded retrieval tool controls one source of leakage, not all knowledge the model might carry.
Ryan Ahern combines leadership-change and sale-exploration signals with AND and OR rules, showing the trade-off between finding acquisitions and raising false alarms. The winning workshop entries then expose a deliberate weakness in the exercise: without a holdout, participants can memorize company names. That failure is the practical lesson. The longer workshop contains this hands-on demonstration and is not a duplicate of the separate short overview.
Watch on YouTube
