John Berryman

Berryman’s work has included GitHub’s code search and early Copilot development, Eventbrite’s discovery systems and the US Patent Office’s search infrastructure. He now runs Arcturus Labs, advising teams on retrieval, prompts, agent design and application evaluation. His search background shapes an emphasis on whether a system returns useful results for an actual task, rather than whether a demonstration looks fluent.[1]

Relevant Search, coauthored with Doug Turnbull, treats relevance as a process of understanding users, inspecting search behavior and adjusting ranking. Its examples use Elasticsearch and Solr features to connect the search engine’s mechanics with business requirements. The approach gives developers a way to debug why useful documents are missing or poorly ranked, then assess whether a change improved retrieval.[4]

In his essay on Jev, Berryman explains how a language model already makes classification-like choices when it decides whether to call a tool and which tool to call. A fixed-choice model can expose those decisions directly, avoiding generation of a longer formatted answer. He argues that established model providers could absorb such capabilities into their existing systems. The mechanism is separate from his competitive forecast about TypeSafe, which remains an argument about incentives and product strategy.[3]