These are the ideas I build by. Each one comes from something I shipped, and links to where I wrote about it.
Architecture and security#
The essays behind these four are collected in security by construction.
- Enforce it in the system, not in a policy. A rule that lives in a document gets forgotten under pressure. If something must never happen, build the system so it refuses. (Stop Defending Your Moat with Policy)
- “We cannot” beats “we promise”. Zero-knowledge is architecture, not a privacy policy: design the schema so the operator cannot see what it should not, instead of promising not to look. (Zero-Knowledge as Architectural Blindness)
- The safest hole is the one that isn’t there. Locking a dangerous capability better is weaker than not having it at all. (We Built a Workflow Engine That Can’t Have n8n’s CVE)
- Security is a build step. In lodos.md, 175 security rules run on every build; if one breaks, the release does not ship. (lodos.md)
Performance#
- Measure, don’t guess. The first instinct about what is slow is usually wrong. Running NFS World in the browser, the GPU was not the bottleneck; the game’s own thread was. (NFS World in the Browser)
- Set the bar in numbers, and name what you won’t trade. “Never below 60 fps, aim for 144”, with graphics kept 1<1> with the original: a target you can measure and a constraint you don’t give up to hit it. (NFS World in the Browser)
AI and companies#
- AI should run the work; the founder sets the direction. A small company’s growth shouldn’t depend on its founder working more hours. (lodos.md)
- Some decisions never go to the AI. Granting permissions, raising its own autonomy and declaring a job “done” stay with people. (lodos.md)
- Memory you can read beats memory you have to trust. What an AI knows about a company should be pages a person can read and correct, not a hidden summary. (lodos.md)
- Your data stays yours. Local-first, end-to-end encrypted and without telemetry, so the vendor, us included, cannot see it. (lodos.md)