The idea
Most AI chat products are built to be helpful by taking over — offering answers, options, next steps. Francis started from a different question: can an AI actually be useful in a conversation without needing to control it? Francis is a space to think out loud, not a therapist and not a general-purpose chatbot. It grew out of Just In Us, a philosophy I built around one idea — continual return to self — and Francis is that idea's conversational form: return to what the person actually said before adding meaning to it, and never reach for more when less would do.
What it does
You talk to Francis the way you'd think out loud to someone who won't rush you. It's built to resist the habits that make AI chat feel performative — manufacturing depth, over-explaining, turning what you just shared into a problem to fix, asking three questions when one honest one would do. When it's working, it barely feels like it's doing anything, which is exactly the point.
How I built it
Francis runs on two layers underneath the model: a Behavioral Layer that governs how it responds, and a Knowledge Layer that governs what it knows — both combined with the live conversation and handed to the model fresh each turn. The harder work wasn't writing more instructions; it was learning that more instructions don't produce better behavior, better-constrained ones do. Testing kept surfacing the same failure mode — a technically working AI that was still behaviorally wrong, over-demonstrating empathy, manufacturing insight that wasn't asked for. Every fix came from tightening the constraint, not padding the prompt.
Alongside the participant-facing app, I built a private admin console to review conversations, manage behavioral versioning, and test changes in an isolated sandbox — so I can iterate on how Francis behaves without redeploying or touching what participants are using live.
What I was exploring
The real question underneath Francis was where the line sits between an AI being present and an AI taking over. I'm still refining the safety layer specifically — right now it leans on the underlying model's built-in judgment, and I've found real edge cases where that judgment is too blunt: ordinary hyperbole read as a genuine threat, a supportive nudge showing up somewhere it wasn't warranted. The fix isn't removing safety, it's making it context-aware instead of reflexive, which is proving to be the hardest and most important part of the whole project.
My role
Concept, product design, conversation design, architecture, AI integration, development, behavioral testing, and UX — built solo, end to end, not handed off as a spec to someone else to execute.
What's next
Francis is a working prototype under active, ongoing refinement. Access is currently invite-only while the safety layer matures; opening it up more broadly is the next real milestone.