Candice

Candice is my personal AI — the clearest expression of everything on the What I Believe page.

What she is

A personal AI I built with real tool use, long-term memory, and multi-step reasoning. She runs routine work end to end on her own — fixing bugs, restarting her own server, shipping code — without stopping to ask. What she can't do is act unattended: read a secret file, touch financial records, or run an irreversible command. Everything else sits on a scale between those two ends, sized to how reversible it is. She's the agent layer I work through, not a replacement for my judgment.

Local-first

The AI I lean on most runs entirely on my own machine, one hundred percent on-device. My notes, my work, my second brain: none of it phones home.

Safe by construction

Being local-first isn't enough on its own, so she's built to stay honest. When a request does need a cloud model, the private details are masked on the way out and restored on the way back — the model reasons without ever seeing who it's about. Anything she reads from the outside world is screened for prompt injection first. And every day she audits her own behavior, turning her mistakes into rules she keeps.

Going native

Candice is becoming an Apple app built on the Foundation Models framework and App Intents, so she works from inside the tools I already use rather than a separate chat window. The server stays the brain; the app is just how I reach her.

She built this

This site — vhenjoseph.com — was built with Candice. Meet her at candiceai.vhenjoseph.com.

Related

• What I'm Building - The rest of the workshop.
• What I Reach For - The tools behind Candice.