TalkToo.ai
Chat with AI characters, learn with AI tutors
Messages exchanged, designed and engineered to feel human, not robotic.
User-created AI characters — each with its own distinct personality and identity.
Building 30 million conversations is a systems problem. Designing them so users want to have them is a product design problem.
Design work on TalkToo covered the full character ecosystem: the creation flow that lets anyone build an AI personality — with its own voice and identity — in under 3 minutes, the conversation UI that adapts across learning, therapy, and casual modes, and the social layer that turns character creation into a community experience. Underneath, a real-time backend handles the 30M+ messages that flow through it, with WebSocket connection pooling, message queuing, and 10+ AI models routed intelligently by conversation context.
Existing AI chatbots had a personality problem — they felt interchangeable. Users could switch models without noticing. The opportunity was to make AI characters feel distinct, memorable, and worth coming back to — a branding problem as much as a product one.
The deeper challenge was context switching: the same user might want a study buddy at 9am, a debate partner at noon, and an emotional support channel at night. Most AI products forced users to choose one mode. TalkToo was designed and built to hold all of them, at scale.
From AI tutors to personal mentors — designed so every conversation feels purposeful.
Character Creation Flow
Most users who want to create an AI character have never written a system prompt. The creation flow abstracts all of that — it asks the user three questions (who is this character, how do they talk, what are they for?) and translates the answers into a working AI personality automatically. Eight different ordering and framing approaches were tested before landing on one where 78% of users completed the full flow without dropping off.
Context-Aware Conversation UI
A conversation with a study tutor looks and feels different from a therapy session — even if both are text-based. A context layer adapts message pacing, visual density, and input affordances based on the character's designated role. Tutoring mode surfaces inline definitions and progress markers. Coaching mode adds reflection prompts and session summaries. The system adapts without asking the user to configure anything.
Social Discovery and Character Sharing
The platform's growth loop depended on character sharing — users discovering each other's creations and adopting them. The discovery surface was designed as a character marketplace with editorial curation, not just a raw list. Each character card surfaces a personality preview, creator info, and a "try it now" interaction that let users sample a character before committing. This single UX decision drove a 3x increase in new character adoption.
Real-Time Messaging at Scale
Handling millions of concurrent conversations required a real-time architecture built on WebSocket connections with intelligent connection pooling, message queuing, and distributed caching for sub-second response times. A microservices approach lets conversation management, AI model routing, and user sessions scale independently based on demand, while a multi-layered notification system — push, email digest, in-app — maintains 99.9% delivery reliability.
Herry has a phenomenal ability to absorb highly complex ideas and translate them into rapid, high-quality prototypes. His focus on fast execution and scalability exceeded all expectations.
Gabe Cowen
CEO, Makr.ai
"Manu has been very persistent and was able to absorb complex ideas and execute in ways that are efficient and scalable. His approach to the backend architecture demonstrated deep technical understanding while maintaining focus on performance. The final product exceeded our expectations in both functionality and design."
Gabe Cowen
CEO, Makr.ai
