AI Engineer
← All speakers

Related work and projects

Bio, Work & Ideas

Ryan Ahern

Conference affiliation: exa

Ryan Ahern was a forward-deployed engineer at Exa in 2026. His open-source work has focused on making web search useful and configurable for AI agents.

In Pydantic AI, Ahern proposed modernizing the Exa integration after seeing a customer issue in which a language model inserted domain filters into its search query. He contributed an implementation that let developers configure those filters directly and recommended returning highlights by default: relevant passages that give agents context with fewer tokens than a full page. His proposal preserved the existing result shape, allowed developers to request full text, and addressed deprecated search options with compatibility in mind.

Ahern also contributed web search text-summary mode and a deferred ExaAgent capability to the Pydantic AI Harness. Exa’s Contents API provides text, highlights, and generated summaries from web pages, giving applications different ways to prepare retrieved material for a model.

Read the topics behind these talks

2 conference talks

▶ Watch ↗

19:12

Version-Controlled Search: The Web as It Was — Ryan Ahern, Exa

Read the full talk →

Key ideas

Scroll to read ↓

Ryan Ahern explains how Exa Snapshot makes past web search available for testing predictive signals—and why a historical cutoff is only part of a sound backtest.

  • Snapshot retrieves captured URL versions or search results at or before a chosen date, making historical web signals available for backtesting.
    6:42 ↗
  • A dated search can keep a later acquisition announcement out of the web result, while other inputs and the agent’s knowledge still need attention.
    10:41 ↗
  • Historical web data can enrich an existing model, supply an agent-generated feature, or support an agent managing the prediction itself.
    11:41 ↗
  • A holdout helps distinguish a reusable signal from a shortcut such as checking a company’s name.
    10:11 ↗
▶ Watch ↗

26:41

Backtest Your Agent Against Yesterday's Web — Ryan Ahern, Exa

Read the full talk →

Key ideas

Scroll to read ↓

Ryan Ahern explains how point-in-time search can test acquisition predictions, then shows why a historical search cutoff still needs a holdout set.

  • A historical search cutoff helps prevent a prediction from reading the announcement of the outcome it is meant to predict.
    9:43 ↗
  • A known URL can be retrieved at or before a chosen date; versioned search can discover relevant pages under a historical cutoff.
    6:12 ↗
  • In the workshop, requiring both leadership turnover and reported sale exploration improved the score relative to accepting either signal, under a scoring rule that penalized false alarms but not misses.
    18:41 ↗
  • Historical retrieval and a holdout solve different evaluation problems: the cutoff limits future web information, while unseen cases test whether a tuned rule generalizes.
    10:42 ↗

References