Inspiration

We were inspired by the common cartoon trope of the angel and the devil on one's shoulder. After experience multiple dilemmas that we faced in real life, we decided to use artificial intelligence simulating an angel and the devil to hopefully gain more insight to help with our decision making.

What it does

The web app allows users to say their burdensome dilemmas into their mics. The question is transcribed automatically into the app and sent to the angel and the devil for their verdict. Finally, their responses are displayed on cards that slide in from the sides. The user is also able to swipe down to see previous questions asked.

How we built it

We made use of the new assistant feature of the OpenAI API with the GPT-4 model. We specified the angel and devil's personalities in the run instructions, allowing the AI model to answer in their respective personas. We also used the browser-built-in web speech API to instantly transcribe the user's voice.

Challenges we ran into

We initially planned on using our own finetuned AI model. It was trained on hundreds of model responses of the 2 personalities and produced more unique and convincing characters compared to GPT-4. However, we soon realised that it took an extremely long time, about 1 min, to generate the responses. This was unacceptable as the user experience would be completely diminished.

Accomplishments that we're proud of

We are extremely proud that we managed to fine-tune a working model in under 24 hours, even though most of the time was spent on the setup of the model. (windows < linux 🤢)

What we learned

We have learnt more about how fine-tuning AI models work and also more about using python as a full-stack implementation of our app.

What's next for Devil's Dilemma

We hope to eventually improve the response time of our model such that we will be able to implement it into our app.

Built With

Share this project:

Updates