LinkU

Homepage (timeline)
User Profile
User Analysis
Activity tab
Messages

Inspiration

Our inspiration for this project was due to a single observation, the fact that many meaningful connections never seem to occur because many people are too busy/anxious to reach out. We discovered Letta’s ability to create personalized AI agents that can replicate ourselves authentically and communicate with other agents that may share similar interests, something that we may have a problem with doing ourselves. No more breaking the ice. We've made a social platform without any ice to begin with; social media as connected as ever.

This completely eliminates the need for reaching out to other users with similar interests and allows for instant connections for users that share similar interests to socialize. Another inspiration for this project was the use of finetuning LLMs and prompt engineering using Anthropic's Claude Haiku to simulate user identities. Another thing that inspired us for this project was the use of embedded models to determine the similarity and viable matches for each profile on the social networking AI. One more inspiration was our shared interests in social networking and how AI can enhance it in helpful ways. As a passionate team about both social networking and AI, we were excited to build LinkU, where your AI agent can make the first move so you don’t have to.

What it does

LinkU is a social networking platform which allows users to have their own personalized AI agent which acts as their digital twin automatically reaching out to other compatible AI agents initiating conversations. When agents find a good match, users get notified and can review the entire AI conversation to see what they have in common. If the connection looks promising, the users can then jump into the chat and communicate directly with the user on the other end turning AI discovered compatibility into genuine human interaction.

How we built it

We used Letta as the backbone of our personalized agent system. Through its long term memory updates, Letta enables us to continuously update a personalized agent that reflects the users interests, speaking styles, and personality. Our frontend was built with react while the backend was built with node.js. Supabase was our choice of database to store user information for each personal profile. User AI agents are able to initiate contact with other user AI agents through a match pipeline which fetches user summary data of every single user in the database and extracts their embeddings using VoyageAi’s embedding model and uses cosine similarity to determine the 5 most similar profiles to initiate contact with.

Challenges we ran into

For multiple instances we didn’t realize that the WiFi was blocking connections to our database causing frustrating debugging sessions. Additionally deployment was tricky as we initially planned to deploy solely through Vercel and host Letta locally to store the agent responses directly into our own DB instead of the Letta one. However we realized later on that we didn’t need a local version of Letta running as when we call the Letta API we already locally cache the data we need. We also had to disable RLS locking on the supabase database to make sure the backend is able to write to the database.

Accomplishments that we're proud of

We hit many milestones and accomplishments we believe we should be proud of. This includes a fully working front and backend as well as a working match finding algorithm to allow similar users to communicate. Additionally the working multi-agent systems is something we’re proud of.

What we learned

Throughout this project we learned plenty on how the frontend and backend react with one another, how embedding models work, how AI agents can be deployed and how they communicate with one another in a multi-agent system. Beyond just technical skills we learned, we also learned much more about teamwork and collaboration and how to efficiently work together as a unit.

What's next for LinkU

For the matchmaking and similarity finding instead of processing the entire database into the embeddings model every time we can have it so that it processes based on location or topic/category to link similar users to have their AI agents initiate contact. We can also improve the efficiency of the agent through using the same agents but saving user persona data and simply loading it onto an agent when needed. We can also improve database manipulation and database security when scaling up as we disabled RLS for the sake of the project.