As three awkward WPI students, we personally know the struggle of social interaction. When tasked with coming up with a response, many struggle to come up with something reasonable quickly. Our product aims to offload the hard thinking to world-class AI, analyzing what was said and giving not just one, but three thoughtfully chosen responses to please anyone wanting to talk.
What it does
The recommender works by first pressing the listen button to turn on the recommender. After you turn on the recommender, it automatically picks up audio and transcribes the latest sentence. Based on the latest sentence, it will give you 3 recommendations as to what to do next.
How we built it
Our application was built using AssemblyAI as our speech to text API, and GPT3/OpenAI as our language model AI to read and create appropriate responses. With both our scripts running in Python, we decided to use Streamlit as our front-end framework to allow streamlined development and integration with our existing scripts, helping us save time and deliver a polished, fully functioning web app.
Challenges we ran into
As 3 exclusively backend software engineers, working the frontend was easily the biggest challenge. Designing and prototyping a working frontend for our application was a nightmare. We went through multiple frameworks and languages to try and see what would work the best with our existing backend code.
Accomplishments that we're proud of
Being able to integrate assemblyAI and GPT3 to solve a hilariously tragic issue.
What we learned
What's next for Recommender
With our current functionality of Recommender, the next steps our team would take to improve this product would be refining the AssemblyAI listening logic and sentence completion detection. We also plan to allow the user to have greater customizability of GPT3, by putting sliders and toggles for various parameters such GPT3 temperature. Allowing the GPT3 to focus on specific settings and tones, such as serious responses for more formal situations, or thematically related responses based on topic would be a great addition for the next evolution of the Recommender.
Log in or sign up for Devpost to join the conversation.