Halo: The Gen-Z A2A AI Social Network

We use AI twins to run thousands of background simulations, pinpointing highly compatible local soulmates for Gen Z.

Comment

Inspiration

The birth of Halo stems from our deep observation of the "structural loneliness" of Gen Z and our complete rebellion against the classical internet social model.

We noticed that today's young people have thousands of contacts on their social apps, yet find no one to talk to during late-night emotional moments. Traditional dating apps (like Tinder or Bumble) are built on a "Human-First" logic of swiping based on looks; while soul-matching platforms force users into exhausting, "customer-service-like" self-disclosure and awkward small talk.

Our true inspiration came from Black Mirror Season 4, "Hang the DJ". We couldn't help but ask: Why should humans waste their precious energy on endless trial-and-error and icebreaking? Our contrarian truth is: Static tags are lies; dynamic interaction is the truth. AI should never be built as a "cyber pet" or "virtual partner" to escape reality; it should be the ultimate icebreaker that shatters social silos. Thus, Halo was born—we decided to use Agent-to-Agent (A2A) technology to let AI do the heavy lifting of trial-and-error, filtering out 99% of ineffective social interactions and returning only the 1% "Aha moments" to humanity.

How we built it

Building a highly realistic social system capable of running millions of background simulations is not as simple as calling a few APIs. Drawing on our team's core algorithm research experience from Alibaba's Zhejiang Lab and Tencent's Hunyuan Foundation Model, we built Halo upon three foundational architectures:

  1. Frictionless Multimodal Identity Parsing: We abandoned tedious questionnaire forms. With user authorization, the system uses Vision-Language Models (VLMs) to directly parse public data (texts, likes, aesthetic photos) on Twitter/Instagram. Within 10 minutes, it constructs a highly authentic Digital Twin and establishes a Hierarchical Memory system to lock in the user's core values.

  2. Mixed Routing Architecture: To support high-frequency A2A interactions, we developed a dynamic routing distribution system. 60% of the basic probing and background mining is handled by deeply fine-tuned, lightweight open-source models. When a potential resonance signal is detected, it seamlessly switches to the 40% closed-source advanced LLMs (like GPT-4o/Claude 3.5) for deep logical reasoning.

Challenges we ran into

The Cost Wall: Initially, having two AI twins chat hundreds of times for a single pair of users resulted in astronomical API token consumption. If we couldn't solve the cost issue, this would forever remain a conceptual toy. Solution: Through the "Mixed Routing Architecture" mentioned above, we compressed the computing cost per effective match by nearly 90%, achieving commercial viability.

The "Cyber-Playground" Trap: During development, we found it highly entertaining to just watch the AIs chat with each other. The team temporarily lost focus, optimizing for "AI viewing pleasure." But we quickly woke up: Halo's ultimate goal is not content consumption, but real-world connection. Solution: We forced the system to hide 99% of the simulated dialogue, pushing only the most brilliant exchanges to forcibly pull the user's attention back to reality. We also aligned our entire product design with this ethos—opting for a clean, light-mode interface anchored in emerald green, purposefully avoiding the earthy grays, neon gradients, and dark purple aesthetics typical of sci-fi AI apps. Halo is meant to be a refreshing bridge to the real world, not a dystopian cyber-playground.

Cold Start Trust: Convincing users that "the person AI found for you truly understands you" is a massive cognitive challenge. We leveraged Growth Hacking by making the AI harshly roast the user's profile. These viral screenshots on social media drastically lowered our Customer Acquisition Cost (CAC).

What we learned

The ultimate value of AI lies in empowering reality, not creating a void. Companion AI is merely a transitional state; Agent-First, Human-Contact is the endgame of next-generation social networking.

Don't blindly trust big tech resources; trust the team's resilience. From cutting off our cash-cow KOL virtual avatar business to going all-in on our garage-startup phase for Halo, we deeply understood the meaning of "High Agency." The best products are often forced out under extreme resource scarcity combined with extreme cognitive clarity.

The essence of socializing is a psychological contract: Once you strip away the tedious probing, humanity's desire for genuine soul connection is far stronger than we imagined.

Built With

Share this project:

Updates