Ideology Detector

Inspiration

This multi-agents system analyzes ideological beliefs through natural conversations. It overcomes the limitations of traditional surveys, which can be easily manipulated.

Papers

"Ideology Detection Using Recursive Neural Networks" by Mohit Iyyer, Peter Enns, Jordan Boyd-Graber, and Philip Resnik, published in the Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (2014).

https://aclanthology.org/P14-1105.pdf

"Beyond Binary Labels: Political Ideology Prediction of Twitter Users" by Daniel Preoţiuc-Pietro, Jordan Carpenter, Christopher Schwartz, Lyle Ungar, and Arthur Spirling, published in the Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (2017).

https://www.sas.upenn.edu/~danielpr/files/moderates17acl.pdf

What it does

Analyzes sentiments through natural conversations that started from a defined set of interview questions, designed to incite opinions in a nonintrusive way.
Continuously learns and adapts from ongoing interactions to generate more insightful and accurate analysis.

How we built it

It uses a dual-AI pipeline.

ElevenLabs AI manages voice interaction and conversation flow.
ChatGPT-4o reads and analyzes the transcript, using the insights we gained about the user to generate system prompts for ElevenLabs AI Agents.

Using multiple agents, we are able to do this in a continuous loop that results in analysis that is more insightful as the interview goes on.

Challenges we ran into

The current version only supports one user at a time, as it is a prototype. A more scalable solution requires more time to implement.
The transcript processing currently takes longer than desired. However, we have already figured out how to do it more efficiently in future versions.
Collaborative iterations on prompt design were challenging. It is also the part that we spent the most development time on.

Accomplishments that we're proud of

Thanks to Lovable, we were able to implement the frontend in a short amount of time. That gave us the luxury of using three.js effects in our project.
With the help of AI tools like Cursor and Lovable, we were able to finish a deliverable product, even though it was our first time using many of the technologies.
The result of the LLM in generating relevant and funny responses surpassed expectations. The analysis is insightful and accurate.

What we learned

How to use multiple AI agents as a team to capitalize on their individual strengths.
The project shows how handcrafted code can complement AI tools to produce unique products in a timely manner.
Deepened understanding of how to iterate prompts and navigate the limitations and personalities of different LLMs. For example, as some models were unable to process our transcripts because of its political nature.