Find Your People

Inspiration

Our personal experiences in therapy, starting companies, ending personal relationships, etc., led us to identify the base-level dynamic of finding the people you work well with as a 'first principles' problem of the human experience, which current social media is addressing in a round-about, inefficient, and ultimately ROI-negative way for the user. When we remove the negative motivation structure of social networks (user-generated public content being the basis for the interest graph), users get all the value with none of the cost.

What it does

Find Your People is a chat interface and multi-agent system that seeks to match users based on their interest/personality profiles.

Users interact with the system in a simple chat-based webapp, while under the hood, unique agents work together to generate user profiles, match users together, and represent their users to other agents. All the users have to do is begin chatting, and they will start seeing matches with other users (and their shared interest criteria) populate a list.

The five agents at work (ConversationAgent, ProfileAgent, MatchAgent, RepresentativeAgent, and RecapAgent) are instances of GPT-4, and are differentiated by their system and few-shot prompts. A communication interface allows the agents to pass messages to each other and work together.

How we built it

The conceptual foundation of Find Your People emerged through extensive dialogue and collaborative iteration within the team. The code that runs the webapp and multi-agent infrastructure was written almost entirely by hand, with minimal use of ChatGPT. The system and few-shot prompts for each agent were initially written by hand, and then iteratively refined using ChatGPT.

Challenges we ran into

Conceptual challenges:

-Distilling the most fundamental ‘first-principles’ aspects of our idea, while simultaneously balancing the limitations of actually implenting that idea during the hackathon (time constraints, the need to demo a product, etc.)

Technical challenges:

-Defining and implementing the multi-agent system and interactivity among the agents.

-Fine-tuning limitations: We initially wished to implement an iterative fine-tuning loop that would train the models on user-agent conversation history, emerging user profiles, etc., but found this unfeasible given the time constraints, and instead decided to leverage system and few-shot prompts as a stopgap.

-Crafting the system and few-shot prompts such that each agent has a clear and specific role, and does not engage in unwanted behaviors.

Accomplishments that we're proud of

-Creating a proof-of-concept in 2.5 days that addresses what we see as a fundamental issue of humanity!

-Writing foundational code for creating and managing interactive multi-agent systems.

-Refining effective system and few-shot prompts for directing agents and reducing unwanted behaviors.

What we learned

-AI agents are notoriously difficult to control! Much care and attention must be put into prompting them well and setting up a solid infrastructure for them to communicate with users and each other.

-Separation of concerns: Just like programming best-practices, we discovered that it was better to avoid having an agent be responsible for performing more than one task. Once the multi-agent infrastructure was set up, it became (relatively) easy to instantiate unique classes of agents and separate concerns among them

What's next for 'Find Your People'

-We need to consult with experts in mental health and behavioral psychology to improve the agents and they way they interact with users and each other.

-We need to take the time necessary to fully implement iterative fine-tuning loops that personalize each agent’s performance based on user data.