Julius Brussee

Brussee’s portfolio dates his Leiden University studies to 2025 onward and describes service on its Data Science and AI education committee. He also works as a founding engineer at Stacklink, an enterprise retrieval platform. Caveman appears in Y Combinator’s active Fall 2026 company directory, which describes the company as an optimization stack for agent workloads.[2][4]

Compression at different points in an agent workflow

Caveman began as instructions that make an agent write terser prose while preserving code and errors. The current project expands that approach to the information an agent reads and the model selected for a task. Its product describes observing workloads, testing changes and retaining improvements that pass evaluation. A reduction in tokens becomes useful only if the task still succeeds, so the comparison needs to account for workload, model and success criteria.[3]

Earlier experiments and learning software

Brussee’s GitHub profile marks Cavekit, Cavemem and Caveman Code as frozen experiments whose ideas moved into Caveman. Revu is a separate application for studying, with a native macOS version built around local storage and spaced repetition, which schedules review as material becomes easier to forget. Auto Build In Public addresses a smaller workflow: it reads a Git history range and produces editable social-post drafts as Markdown files. The output remains available for review before publication.[1][5][6]