Software engineer in Berlin, currently at BeeHealthy. Most of what I build lives in the gap between developers and their AI agents — making the context smaller, the feedback tighter, and the tooling honest about what it actually did.

I write about it at pecan.si — agents, Rust, open source, and the occasional detour.

Project	★
tokf		Declarative TOML pipelines that compress CLI output before it reaches an LLM context — 60–90% fewer tokens. tokf.net
rable		A from-scratch GNU Bash 5.3 parser in Rust, ~10× faster than Parable, with PyO3 bindings. The parsing backend in tokf.
upsert		Spring Data JPA extension adding native upsert operations for PostgreSQL and MySQL.
cargo-dupes		Cargo subcommand that finds duplicate and near-duplicate code via AST normalisation.
cargo-lint-extra		Configurable Rust lints beyond rustfmt and Clippy.
share-the-mark		Browser extension that turns an annotated live page into a Markdown brief for a coding agent.

• Agent tooling — context compression, routing, and measurement: ccroute, boarder, mcp-coder-bench, rusted-claude-meter
• Rust — CLIs, parsers, and small sharp tools
• JVM & Kotlin — Spring Data, Gradle plugins, coverage and build tooling
• Developer experience — the unglamorous parts: builds, caches, linting, CI

_{Colours and the peanut borrowed from my own site, Mixed Nuts.}