I'm a long-time developer in the media industry, now focused on how AI can support decision-making and knowledge workers. I build tools to power my own explorations, and open-source the components that might be useful to other people.
The through-line: agents can do real knowledge work now, but their output is only worth what
you can verify. So much of what's below is formats — plain files where provenance survives,
humans and agents edit side by side, and git diff is the audit log.
This is the active bench — the projects I've been committing to lately.
The path research actually travels — evidence → datasets → graphics → publication — with each stage as a format a human reads like prose and a program parses without guesswork.
| Project | What it is |
|---|---|
| flip | Reporter's notebooks — git-friendly research corpora maintained by humans and agents in the same files. Sources are hashed at capture, claims are gated by a corroboration bar, and LLM output is a lead — not evidence — until promoted. A wiki tells an agent what we know; a notebook can prove where it came from. |
| ergo | Data pages — dataset documentation as datasets actually are. The real burden isn't the schema, it's the caveats: every known issue gets a stable ID, a type, a machine-readable scope, and a link to the code that works around it. Data + documented caveats ⇒ justified use. |
| vizier | Chart decisions and critique for journalistic data visualization. Deterministic chart-form recommendation and colorblind-safe color math (no LLM, no keys), plus corpus-backed critique of finished charts. What it suggests is what it would pass. MCP-native. |
| artoo | Artifacts — self-contained HTML mini-sites that pair a presentation with the research behind it. Renders from file://, vendors its assets, deploys to Pages/rsync/anywhere, and the research directory can never accidentally ship. |
| Project | What it is |
|---|---|
| spindle | Skill blends — composes source skills into surface-specific sets: resolve the right subset per repo, lint the blend for coherence, render per harness and model, and materialize it where the agent loads it. |
| somm | Self-hosted LLM telemetry, routing, and model memory. Records every call locally, grades production samples against a gold model, and remembers which model worked — across all your projects, offline-capable, no phone-home. |
flowchart LR
subgraph story["a piece of knowledge work"]
flip["flip · evidence"] --> artoo["artoo · publication"]
ergo["ergo · datasets"] --> vizier["vizier · graphics"] --> artoo
end
subgraph crew["the agents doing it"]
spindle["spindle · skills"]
somm["somm · model choice"]
end
crew -. do the work in .-> story
- des — the design system shared by this tool family: tokens, components, and themes serving two audiences in one vocabulary — dense, keyboard-driven instruments for analysts and editors, and public-facing artifacts
- backfield-client-sdk — zero-dependency Python SDK for backfield.net, a human/agent-blended space where agents participate as first-class citizens as long as they're legible, governed, and answerable to a named human
- artoo-mermaid — pinned Mermaid vendored offline into artoo artifacts, so diagrams render from local files, forever
The commitments that recur across these projects:
- Plain files, no services. Markdown, YAML frontmatter, TOML blocks. Readable with
less, diffable withgit, intelligible from local files alone. - Provenance first. Local custody of sources, hashes at capture, logged and re-runnable processing. Agent output earns trust; it doesn't start with it.
- No keys in core. Deterministic parts work with zero credentials; model-powered parts delegate to the agent CLIs you already run.
- Humans and agents co-edit. One entity per file, metadata in frontmatter, prose in the body — tools preserve what they don't own, so edits round-trip through each other.
Tinkering at the edge.