Fred = AI Curator

Inspiration

Our inspiration was various research papers on recommendation systems leveraging LLMs as priors on human behavior. We asked: what if the LLM made recommendations based on your interests directly using natural language?

What it does

Fred works by using a Large Language Model (LLM) to curate and rank content based on user feedback, adapting in real-time to evolving preferences. Fred manages a multi-armed bandit algorithm to balance showing familiar vs. new content, considering user interactions and time of day.

How we built it

We used Streamlit for the front-end. Chat is handled with Langchain and Mistral API. A simple multi-armed bandits algorithm is configured by the LLM using it's past history of interacting with the user and prior knowledge. We leverage a number of open source APIs to fetch data from Google News, Arxiv and more.

Challenges we ran into

A team mate left the team early on, so it was just two of us.

Accomplishments that we're proud of

We'd never used Streamlit before but fortunately managed to build a working UI. Barely every user pydantic -- we like structured generation.

What we learned

Mistral Large 2 can manage structured data like large JSON with ease!

What's next for Fred = AI Curator

If there is interest in the project, we'll try scaling it up to a small number of users at a local university and add newsletters plus more sources of research papers and news.

Built With

• langchain
• mistral
• pydantic
• python
• streamlit