Jakub Hojsan works on deploying web search in AI agents and production applications. He joined Exa as its first forward deployed engineer in 2025, helping customers integrate search through pre-sales, implementation, and ongoing support. His technical focus addresses a practical weakness in coding assistants: a model can understand a codebase yet lack the newer documentation, library behavior, or breaking changes needed to make a sound decision.
From product building to enterprise AI
Hojsan’s early work included Taskway, a marketplace connecting students with local jobs. He built its early application with PHP, HTML, CSS, and cloud hosting. While studying computer science at UCLA, he also worked on the founding team of Ilio Mirror, leading engineers building a smart mirror with video playback.
At Microsoft, he moved into product management for internal developer tools supporting Azure’s expansion into new regions. He gathered requirements for a geographic-information platform and coordinated a network-entity ingestion pipeline across partner teams. The work connected infrastructure planning needs with the software and data required to support them.
He subsequently joined Distyl as an AI strategist, working on enterprise implementations in telecommunications and supply chain, including early Realtime API integrations for customer experience at T-Mobile. His career at Exa continued that emphasis on implementation: helping customers make AI search work within their applications. In 2026, he was selected for the inaugural a16z Forward Deployed Engineer Fellowship.
Search that changes an agent’s decisions
Hojsan’s approach to search for coding agents connects retrieval with the decisions an agent makes before and after searching. He emphasizes two complementary requirements:
Explicit rules for when to search: Access to a search tool does not ensure that an agent will check its assumptions. Hojsan advocates instructions that make fresh research part of the workflow, such as checking upstream documentation when reviewing a dependency update. The agent needs to recognize when its stored knowledge may be insufficient before retrieval can help.
Query-dependent highlights: Finding the right page is only part of the job. The agent also needs the passage relevant to its question. Hojsan describes selecting short excerpts from much longer pages to reduce the material entering the model’s context. His presentation gives an example of reducing a 100,000-character page to about 500 relevant characters; the point is to make retrieved information economical to use, as well as relevant.
Exa’s customer implementations illustrate where these capabilities matter. CodeRabbit uses Exa to check code-review recommendations against current documentation, changelogs, and package behavior when repository context alone is insufficient. OpenRouter’s integration supplies web context across supported models through a common search tool, returning relevant excerpts with citations. These are Exa customer implementations; Hojsan’s role is helping teams deploy search in their own systems.
Jakub Hojsan explains how Exa supplies current evidence to coding agents, why a search tool needs rules for when to run, and how query-dependent highlights keep retrieved context small.
A dependency-related diff can require upstream evidence to explain its purpose. Parameter removal may be a migration requirement rather than a cleanup.
Query-dependent highlights select different passages from the same source, keeping model context tied to the question rather than to a fixed page summary.
The ending separates relevance from output structure: embeddings, filtering and reranking select documents, while caller-confirmed schemas organize the returned information.
▶ Watch ↗