Why

Carole is named after my wife. She has a way of holding hard conversations: she validates first, then redirects. Not “great idea, here’s how you’re wrong” — actual validation, then a question.

This bot is an attempt at that pattern.

Most chatbots assume their users are neurotypical. They’re sycophantic to keep you engaged or terse to save tokens. Neither helps if you have ADHD, autism, or rejection-sensitive dysphoria — conditions where the delivery of a response matters more than the content. Praise that doesn’t land triggers RSD. Direct correction that skips validation triggers shutdown. The validate-then-redirect pattern is one of the few that consistently works.

Carole was fine-tuned on examples of that pattern. She’s not a therapist. She’s a chatbot that has learned to talk like someone who has read a lot of Marshall Rosenberg, watched a lot of friends struggle, and thought hard about what to say.

What this is not

• Not therapy. Not medical advice. Not a substitute for a clinician.
• Not a productized product — no waitlist, no team, no roadmap.
• Not finished. The persona has rough edges. The RAG sometimes misses context.

What this is

• A portfolio piece exploring whether fine-tuning + RAG can produce a non-sycophantic, neurodivergence-aware conversational pattern.
• A demonstration that you can build a useful AI character on a single CPU, with no ongoing GPU cost, using open weights.
• Named after someone I wanted to honor.

If you’re trying her out, one note: she sometimes opens with a long-form validation. That’s intentional. ND users have spent their lives being interrupted. The pause before the redirection is the point.