I ship complete software with AI as the co-pilot and my judgement at the wheel: local-first, security-hardened, measured before it ships.
ScrollData lives on the user's disk. Sync is a feature, not a landlord, so nothing I build needs my server to keep working.
If it isn't evaluated, it isn't done. The AI triage only moved to model routing after it beat the rules baseline on a 40-case eval set: 88% against 70%.
Every dependency and open surface has to earn its place. Protec went through a six-area security audit and shipped 16 hardening fixes.
I build with AI tools and keep the decisions, the review and the testing. If it ships, I can explain why it works and how it fails.


