Each area is tied to software we have shipped or a note we have published.
01
Tools that teach without risk
How do you let a beginner do the real, dangerous thing and learn from it, without paying for the mistake? Guards that refuse a command and explain why, practice environments that are real enough to matter, and progress that comes from doing rather than reading.
AI that drafts and people who decide. Where an assistant should act on its own, where it must stop and hand over, how it learns a person’s voice from their own edits, and how to make each of those boundaries visible to the person relying on it.
Software whose data stays with the person or business it belongs to: a private memory per customer rather than a shared pool, and a choice of AI provider, including models that run locally when the data has to stay put.
Internal tools, admin dashboards and small web apps rarely get a security review, and they often hold the most sensitive view of a business. We study the failure modes that pass review because each step looks correct, and the designs that make them impossible to write.