YaYa: A Relationship-Grown AI Companion

Inspiration

Most AI companions today are selected from preset roles, but they often feel generic and scripted. We wanted to build something more personal: a virtual human that grows from the user’s own relationship data instead of being chosen from a menu.

That idea became YaYa. Users can import relationship-style conversation data, connect external channels, and generate a personalized virtual human that feels familiar, supportive, and emotionally grounded.

What the application does

YaYa is a relationship-native virtual human interface.

It lets users:

• import conversation data such as Discord or WeChat-style excerpts
• analyze relationship tone, care patterns, and language habits
• generate one personalized virtual human
• enter a live conversation mode with that virtual human
• connect OpenClaw as the execution and channel shell

The product is designed as a two-screen experience:

1. Setup / Generation

   • import source data
   • connect OpenClaw
   • generate the virtual human

2. Live Session

   • left side: the virtual human
   • right side: the chat interface
   • the system supports voice, expression, and ambience layers around the conversation

How it was built

We separated the system into clear layers.

• MiniMax M2.7 powers relational analysis, persona generation, dialogue, and task intent
• Gemini TTS powers the speech layer
• Imagen 4 powers the visual avatar layer
• Lyria 3 Clip powers ambient emotional audio
• OpenClaw handles channels, runtime shell, and execution

We also built a standalone YaYa backend, a two-screen web interface, and a local OpenClaw bridge so that real-time Discord-style messages can be routed into the YaYa system.

Challenges we ran into

One major challenge was drawing the right boundary between YaYa and OpenClaw.

We decided that:

• historical relationship data should belong to YaYa
• real-time channels and execution should belong to OpenClaw

This gave us a cleaner architecture:

• relationship history -> YaYa
• real-time channel + execution -> OpenClaw -> YaYa backend

Another challenge was making the product feel like a live conversation instead of an AI dashboard. That led us to redesign the UI into a setup screen and a conversation-first screen.

What makes it different

YaYa is not just a chatbot or a preset roleplay system.

It is designed as a relationship-grown virtual human:

• generated from user-imported relational data
• expressed through voice, avatar, and ambience
• connected to external channels through OpenClaw
• focused on entering a live session quickly, not staying in a dashboard

What we are proud of

We are proud that YaYa already works as a real product direction, not just a concept.

We built:

• a working M2.7-powered backend
• a setup-to-live-session product flow
• a local OpenClaw bridge
• a conversation-first interface
• a clean architecture boundary between data generation and real-time execution

What we learned

We learned that building a convincing AI companion is not just about using stronger models. It is about making the product feel coherent.

The most important design decision was separating:

• relationship generation
• real-time conversation
• execution channels
• visual and audio expression

That separation made the whole system much clearer.

Next steps

Our next steps are:

• improve automated import flows for more real chat history sources
• strengthen the live visual expression layer
• deepen long-term memory and personalization
• expand real-time OpenClaw channel integrations
• make the virtual human feel even more alive turn by turn

Built With

• gemini
• gemini-tts
• imagen-4
• lyria-3-clip
• minimax
• next.js
• openclaw
• typescript