I build public, versioned work on governed, auditable agentic-AI systems.
The core idea is simple: agents are not just prompts you tune. They are systems made of artifacts you can design, version, inspect, test, and govern. Things like manifests, prompts, skills, tool specs, memory and state strategies, schemas, evals, policies, and runtime configs.
My background is in clinical data and regulated environments, where "it works in the demo" was never good enough. That shaped how I think about AI. The interesting question isn't how you talk to a model. It's what you build around it.
Can you version it?
Can you inspect it?
Can you govern it?
agentic-ai-artifact-taxonomy
A framework-neutral taxonomy for the artifacts agentic systems are actually made of. Shared vocabulary over another clever demo. This is the source of truth the rest of my work points back to.
agentic-artifact-builder (try it live)
A small browser app for working with the taxonomy: pick an artifact type, fill in guided fields, and generate clean, public-safe starter files. No install.
agent-librarian
A deterministic CLI that scans an agent repo, catalogs the artifacts it finds, and validates them. The point is making a pile of files reviewable.
journal-agent
A reference implementation of the taxonomy with a clean public/private split: control-plane artifacts stay public and inspectable, private data stays private.
- Governed > clever. An AI you can't audit isn't a system. It's a liability with good manners.
- Artifacts, not vibes. If it matters, it should be addressable, versionable, inspectable, and governable.
- Boundaries by design. Decide what's public and what's private before you build, not after.