One idea — say more with fewer tokens — compounded into five open-source products: a skill, a workflow, a memory layer, a terminal agent, and a fine-tuned model.
Everyone's racing to build better AI models. I build the tooling around them. Caveman, my Claude Code skill that strips ~65% of an agent's tokens without losing accuracy, hit #1 on Hacker News and crossed 70,000 GitHub stars. It now anchors a pack of seven agent skills I ship as one npm install.
I'm also a founding engineer at Stacklink, an enterprise RAG platform, and I'm building Revu, a macOS study app that brings spaced repetition and multi-agent AI to studying. Both share an instinct I keep coming back to: build things that get better the more you use them. Stacklink's search sharpens with every document you give it; Revu learns what you actually forget and schedules around it.
I study Data Science & AI at Leiden University, where I sit on the Education Committee and help shape the curriculum. I write about building AI systems and shipping indie software on Polder, my Substack.
BSc
2025 — Present
Leiden University — LIACS
Caveman is at 99,984 GitHub stars and approaching 100,000 — the full ecosystem crossed 102,000 across caveman, cavekit, cavemem, caveman-code, and cavegemma. Elasticsearch Labs shipped elastic-caveman this week, showing 64% token reduction across eight real Elasticsearch scenarios. Maintaining the codebase, triaging issues, and shipping new modes. Heads-down on Stacklink: building the incremental sync pipeline and hybrid retrieval layer for the TU Eindhoven Innovation Contest. Cavekit, Revu, and Polder posts in parallel.
Last updated 2026-08-21
Broadcast · RTL · Editie NL · May 2026
RTL's Editie NL covered the Leiden student behind Caveman and the “primitive language” trick that trims AI token use by ~65%.