Papersync

Inspiration

As students, we are constantly bombarded by the jargon that is passed around in academic literature. The endless searching in citations that only grows the further you go down the rabbit hole only makes it harder for us to learn. Our struggle is not unfounded because academic literature is meant to be this way; it allows the best of best to communicate ideas in an efficient and effective way to push the frontiers of science. However, dense academic publications to the average person only overwhelms them and discourages them to pursue science. We believe in order to push the frontiers of science, we simultaneously need an efficient system of communicate but also a gateway for the inexperienced to be inspired, so that they may grow to one day contribute to the world themselves.

What it does

Papersync is a journal. This means that there are researchers who post content to the platform, and users who consume it. There are a few key features that differentiate Papersync, however.

Firstly, we have an interest algorithm that utilizes the semantic power of GPT-4 to match a person's interests to the vast corpus of academic papers out there, so that they are able to read what interests them. This solves a central problem with academic literature: journals are often specialized and unique. General journals like Arxiv and Google Scholar are often hard to parse through and often discourage beginners because of an unclear starting point.

Secondly, we have a retention rate algorithm that is based on a user's retention and likes, and is skewed to value more inexperienced people. What this means is that the content that a researcher posts would be promoted on the site more if they explained with more clarity and promotes less use of academic jargon.

Lastly, we detect fraud using GPT-4's semantic capabilities to keep the platform clean and devoid of irrelevant or harmful content. This filter before content is posted is meant to be a loose form of peer review. We believe as the capabilities of Large Language Models grow, the possibility of automated peer review could eventually become a reality, which would allow us get to the truth faster and allow us to progress science more effectively.

We believe that the combination of these features (along with a form of monetization from advertisements or partnerships with academic institutions) will incentivize researchers to share their knowledge on the platform, ultimately leading to a more inclusive scientific community and leading to future scientists and more discovery.

How we built it

We used Next.js, a framework for React. We used supabase as our backend, and made heavy use of GPT-4 for many of our features.

Challenges we ran into

We had some trouble getting an pdf storage feature to work. Eventually, we scrapped the idea and just went with storing the link of the pdf. Images were also something that we had some trouble working with. Eventually we got all the links to work. We had some trouble with getting the API calls to work, because the message data needed to be formatted in a very specific way.

Accomplishments that we're proud of

We're really proud that we got the version control to work and that we did not run into to many problems with merges, with so many different parts moving and people working on different parts of the repo.

What we learned

We experimented with the prompt tuning and recommendation algorithm using GPT-4, and learned a lot about how we cold use it to both filter content as well as how we could use it with embeddings.