RaiTalk

Talk to your AI psychologist. RaiTalk sat where technology, psychology, and cognitive science meet: how emotional and rational thinking interact, and what an AI looks like when it surfaces that interaction instead of papering over it.

Status: Sunset (Jul 2025 – May 2026). The product is no longer being actively developed, but the insights it produced carried directly into KOS, the exocortex project that followed it.

The Problem

Someone comes to a chatbot in the middle of a bad week and gets sympathy. That’s it. No real understanding, no help seeing the pattern underneath it — just a well-phrased “that sounds hard.”

Therapy works differently. It takes sessions before a pattern even becomes visible to the person living it. Most people never get that far.

RaiTalk was my attempt to close that gap — not by replacing therapy, but by helping someone see their own pattern faster than a conversation usually allows.

The Vision

I didn’t want to build something that gives advice. Advice is cheap, and it rarely changes anything. I wanted to build something that asks the right question at the right moment, and lets the user arrive at their own answer — because the realizations people actually trust are the ones they reach themselves.

Key Product Decisions

A few decisions shaped how it worked.

Reflection over advice. Every time the system was tempted to tell the user what to think, I pulled it back to asking instead. People open up more when they’re not being managed.

Conversation over questionnaires. Structure up front kills honesty. A form makes people perform; a conversation lets them slip.

Session memory, not a permanent profile. I didn’t want RaiTalk to feel like it was building a file on someone. Privacy wasn’t a feature here — it was the reason people would say the true thing instead of the safe thing.

The Engineering

Underneath the conversation: a layer that maps what someone says to known psychological patterns, tracks how their mood moves within a single session, and uses both to decide what to ask next — not to sound clever, but to keep steering toward the moment where the user says the true thing out loud.

What Actually Happened

It went from idea to production in 2 weeks. People used it, in real conversations, to work through real things — 20% of the early cohort came back without paid acquisition — and the same handful of cognitive patterns kept showing up across completely different people, which told me the idea wasn’t imagined.

What I didn’t expect: model sophistication mattered less than I thought it would. The moments that actually helped someone came from product judgment — knowing when to push, when to stay quiet — not from a better model. I’m still not sure how to teach that to a system. That’s roughly where KOS picks up.

---

If you work at the intersection of technology, psychology, and cognitive science, I’d like to hear from you — let’s connect or book time on Topmate.