Inspiration
Inspired by the CASM Blue sky misinformation detection theme, we wanted to create a project that both achieves the theme and is able to do so in real time.
What it does
Our app takes real-time data from bluesky (better twitter) posts and analyses them using BERT keyword extraction. It then uses a much more complex gemini model to detect prime misinformative message suspects. The model uses google search and sources from this search to back up its claims. Users can also use the dashboard to filter by certain keywords or themes to further track down misinformative groups or spam.
How we built it
The website is built with a ReactJS frontend, with authentication done by Clerk. The backend uses mongodb as a databse and Python for LLM and BERT inference using FastAPI. Real time communication is made possible through WebSockets.
Challenges we ran into
Whilst setting up WebSockets was initially simple, getting it to scale with the huge traffic of bluesky posts was a major challenge, and required a huge amount of batching and optimisation. We also struggled with best utilising BERT and Gemini to accurately track down misinformation, since at times even a set of Google searches will not guarantee an accurate response.
Accomplishments that we're proud of
At the end of the 24 hours, we were able to achieve real time BlueSky post monitoring, with good accuracy in detecting potentially misinformative or inflammatory messages. We were able to integrate this with a sleek user dashboard interface, with a variety of different ways to meld the incoming data to the User's needs.
What we learned
We learnt a great deal about LLMs and its applications in high throughput data. We also learnt how to better make use of WebSockets in a much higher stress environment, where performance is necessary. Lastly, and most importantly, we made some very important lessons in coordination and teamwork ⭐
What's next for Blue Sentry
We plan to make Blue Sentry more streamlined and performant as we can, and improve its usability for professionals and data enthusiasts worldwide! Our aim is to craft an app with the performance and flexibility of something like WireShark, but for tweets and BlueSky posts!
Log in or sign up for Devpost to join the conversation.