Jeffrey Wang

Jeffrey Wang is co-founder of Exa, which builds web search infrastructure for AI applications and agents. Alongside that search work, he has developed an engineering approach to go-to-market: maintain current information about customers, give software access to it, and build tools that help people act on it.

From college search experiments to Exa

Wang studied computer science and philosophy at Harvard, where he ran a GPU cluster in his dorm room and was roommates with his future co-founder, Will Bryk. The two collaborated on early search-engine experiments in college. Before Exa, Wang spent three years building data and web infrastructure at Plaid.

Their search company initially operated as Metaphor. It pursued retrieval by meaning: a descriptive request could find relevant webpages even when they did not contain the same words. A search for climate-related startups, for example, could return company websites rather than articles listing those companies. In early 2023, the team shifted toward a search API for AI applications; Metaphor became Exa in January 2024.

Exa developed its own crawling infrastructure, indexes, embedding models, and retrieval systems to collect web information, identify relevant pages, and deliver their contents to applications. Wang describes the underlying approach as embeddings over the internet, enabling semantic searches that identify and classify companies or retrieve information for agents. Customers include Cursor and Cognition. In May 2026, Exa announced a $250 million Series C led by Andreessen Horowitz to expand model training and search infrastructure.

Building a go-to-market system

Wang’s interest in customer-facing systems grew out of a mistake he acknowledges: the founders’ early emphasis on engineering left them doing too little sales and marketing. He argues that a company needs both a good product and the ability to get it into people’s hands. When Exa launched in mid-2023, GPT-4 had made it possible to automate substantial parts of customer research and sales work, and he began approaching go-to-market through agents and the data they need.

• Customer dashboards: In his 2026 account of Exa’s internal systems, Wang explains what this means in practice. The company’s ideal-customer-profile dashboard uses Exa search to classify potential customers into segments such as model providers, AI coding platforms, and go-to-market intelligence tools. Users can move from a market-wide view to detailed company information, including estimates of potential spending. Request Lens adds signals from actual product use: new signups, sharp usage increases, customers who stop searching, and activity from accounts the team particularly wants to reach. These systems connect external market research with internal customer behavior.
• Agent access and customizability: Agents make that information usable beyond the dashboards. Exa’s go-to-market team uses agents in Slack to investigate accounts and build customer demos. Wang’s prerequisite is programmatic access to internal and external data, whether through an API, a command-line interface, or the Model Context Protocol. He applies the same test to purchased software: can the team adapt it to its work and give agents useful access? Exa uses Salesforce because its established sales data model saves the team from making those design choices itself, while its MCP interface lets agents work with the data. For Wang, customizability matters more than whether software is bought or built.
• Familiar interfaces: He also gives consistent interfaces a continuing role. A generated interface can answer an unusual question, while a familiar dashboard lets people learn controls and return to recurring work. His approach combines those stable views with conversational agents that can investigate questions beyond the dashboard’s predefined options.

Jeffbot: modeling judgment and limiting permissions

Jeffbot is Wang’s more personal experiment in this approach. He built the assistant to approximate his writing and judgment, analyzing roughly 760 of his emails to characterize his voice and hundreds of past decisions recorded in Slack and email to create evaluations. Those evaluations helped calibrate the agent’s responses against decisions he had actually made. Colleagues use it to draft emails, Slack messages, and proposed answers with company context.

Jeffbot’s permissions depend on who invokes it. Wang’s use can include reads and writes across systems he can access. Other employees receive a restricted set of tools and can use it to draft messages; they do not inherit his full ability to act. That distinction is central to the project: an assistant modeled on a founder’s judgment need not give every user the founder’s privileges.

Engineering around customer needs

At the time of the talk, Exa’s forward deployed engineers supported customer deals while maintaining and extending internal AI systems. Wang sees a practical reason to combine those jobs: engineers who encounter friction in customer work can also build tools that remove it for themselves and account executives. Other sales staff receive training to use the systems effectively without being expected to build every interface. He presents this as a working organizational model, while acknowledging that the responsibilities may become more specialized as the team grows.

Exa’s search work also includes the research team’s WebCode evaluation suite, which tests search for coding agents by separating retrieval quality from content extraction and using tasks that cannot be adequately answered from model training memory. It addresses the same practical concern that runs through Wang’s internal systems: capable models need useful information and usable interfaces before they can do useful work.