The workshop

Projects

A handful of things built slowly and kept alive — some growing, some in the field, one that is this very garden.

Corpus

The quiet engine behind this garden — essays, projects and library shelves compiled from plain text into a fast, static, self-hosted site.

This website is a project too, and it is dogfooded daily. Corpus compiles a folder of Markdown into a static garden: writing with a proper table of contents, projects with field notes, a library of books and quotes, and a command palette to wander by keyboard. The aesthetic constraint is the technical one — everything must survive a move to a new machine as a folder of text files.

Marginalia

A reading environment for serious readers — highlights, marginal notes and slow revisits, kept as plain text and never locked in.

Books deserve better margins than a cloud service that might vanish. Marginalia is an experiment in reading software that respects its material: annotations stored as Markdown, quotes linked back to their source, and a reading interface designed for length rather than speed. It is the sibling project of this garden — one reads books, the other grows essays.

Personal Knowledge OS

An AI-native memory system that remembers the way the mind forgets — associatively, by salience, never in folders.

The working thesis: a personal knowledge system fails when it treats notes like files and succeeds when it treats them like memories. Personal Knowledge OS ingests writing, code and reading, then organizes them by association and decay rather than hierarchy. Retrieval is conversational; maintenance is a matter of pruning, not filing.

Principles, so far:

Salience over taxonomy. If a note is never revisited, it should fade instead of clutter.
The graph is the interface. Folders are a fallback, never a home.
Plain text at rest. Everything is legible without the system that reads it.

The current build is a Rust core with a Python research layer and small language models handling the associative indexing. It is the slowest, most rewarding project on this page.

Observatory

Instrumentation for AI agents in production — traces, evals and the boring telemetry that makes autonomy trustworthy.

Agents fail quietly. Observatory gives them a telescope: end-to-end traces across model calls and tools, continuous evaluation against a slowly curated suite, and alerting tuned to be boring — because the best signal from an observability system is a calm dashboard. The hard problem is not collecting events. It is deciding, in advance, what a failure looks like.