Make claims auditable.
Paths, links, diffs, tests. If it matters, it should be checkable.
Full throttle. Tight reins.
Build AI systems that ship.
AI workflow design, tooling, and evaluation for teams that ship. Public work, documented methods, and working examples on Projects and GitHub.
Old software instincts miss these failure modes.
Sycophancy spirals, generation addiction, phantom capability. New traps need new patterns.
Stop shipping feelings.
If the spec is vibes, the result is vibes.
Blast radius is real.
Confirmation gates and rollback paths beat good intentions after the fact.
Know when to stop.
Timeboxes and ship/stop signals beat endless polish loops.
How I work
Clear evidence, explicit stop points, and rollback paths where the risk is real.
If you are new here
Pick the surface that matches how you prefer to evaluate the work.
Start with evidence
Proof strip
A quick read on what is already public, inspectable, and tied to real operating claims.
Latest note
The Receipt Was Lying
March 24, 2026. Compound-command bypass, VDP rejection, and what it exposed about LLM triage failure.
Public work
Projects you can inspect
Freeverse, Alcove, and Golden Thread are public. Private tracks stay labeled as private.
Method library
Patterns with failure modes
The methods are written as repeatable operating rules, with anti-patterns sitting beside them.
Open surface
GitHub is part of the proof
Public repos, linked capabilities documents, and build artifacts stay inspectable instead of hand-waved.
What I do
Mechanisms, not magic.
Workflow design
Clear loops, clear handoffs, and decision points that hold up under pressure.
Skill/tooling build
Automations that don't rot in a week.
Evaluation
Adopt the proven. Test the promising. Ignore the noise.
Start with something concrete.
I reply faster to structured briefs. Two ways to start: