Guzhi

Landing
knowledge graph

Inspiration Traditional dating and social apps cause burnout with endless swiping and superficial judgments. Guzhi solves this by using AI avatars to explore social chemistry and compatibility before users ever start chatting.

What it does Guzhi creates an AI digital twin reflecting your personality, values, and humor, then places it in a simulated virtual town. The agents interact autonomously, evaluate conversational rapport, and generate personalized match reports with icebreakers for real-life connections.

How we built it

Multi-agent orchestration layer driven by LLMs for memory, planning, and dialogue.
Vector database architecture for long-term episodic and semantic memory.
Interactive 2D/3D web and mobile interface for live simulation viewing.
Asynchronous backend services handling simulation ticks and compatibility scoring.

Challenges we ran into

Preventing persona drift to keep agent behavior consistent with the user's authentic personality.
Optimizing token usage and latency across concurrent, multi-turn agent interactions.
Designing an objective scoring model to evaluate subjective emotional chemistry.

Accomplishments that we're proud of

Built a functional virtual sandbox with autonomous, emergent agent conversations.
Designed a smooth transition from simulated compatibility tests to human-to-human introductions.
Created an intuitive interface balancing game-like exploration with genuine connection.

What we learned

Multi-agent simulations uncover deep conversational rapport that static profile tags miss.
Accurate avatars depend on how a person reacts and thinks, not just lists of hobbies.
Users prefer transparent match reasoning over black-box percentage scores.

What's next for Guzhi

Real-time collaborative scenarios (such as co-op games or trip planning) for user pairs.
Granular privacy controls over memory and persona boundaries.
Multimodal upgrades including voice synthesis and expressive avatar gestures.