I am working on AI systems that keep contact with the thing they claim to understand.
Current work
The focus is agent infrastructure and physical AI. The shared pattern is simple: preserve the artifact, stabilize object identity, expose typed relations, keep hard checks outside the model, and learn only the residual ambiguity.
That pattern covers mechanical fixtures, CAD and DFM automation, tactical behavior graphs, agent harnesses, and the DXF-to-object/GNN plan.
Interests
- Physical AI and CAD-native representation
- Editable geometry, stable objects, typed relation graphs, deterministic checks, and learned repair only where the geometry remains ambiguous.
- Agent infrastructure
- Thin harnesses, explicit state, provenance, eval loops, and interfaces that improve as base models improve.
- Cognitive security
- Calibrated action under shift: evidence-bearing inference, analyst-legible drilldown, and confidence that does not pretend context is stable.
- Writing and graph memory
- Short public pages, dense essays, marginalia, and graph views of the archive underneath.
- Music as method
- Motif, variation, timing, error recovery, and the discipline of keeping expression inside a form sturdy enough to carry it.
Prose
I want the public site to stay simple while the writing keeps its strangeness. The practical work and the mystical material should make each other clearer.
The site should introduce the work, the music, the engineering habits, and the older essays without flattening them into one tone.
Excited about designing simple affordances within complex environments.