Trialogue Labs - AI-Powered Conversation Moderation

For the last 8 months or so, I've been writing a webapp for Dr. Greg Wurm, in the Sociology department. Initially, it was supposed to be a refactor of his vibe-coded prototype, but there was so much work and retooling to be done that we decided to scrap the code entirely. This project would be a continuation of that process.

This project will eventually be used for real-world research and possibly a startup.

Description

The idea at the core of Trialogue is to use an LLM to serve as a mediator/moderator between two users (possibly more, but that's a later feature). In addition to the Moderator, each user also has access to a private AI coach that can help them draft responses to difficult. We see potential uses for couple's therapy, political debates, class discussions, and difficult personal conversations with loved ones, but those are certainly not the only options.

The goal of this project is to complete work on the current "generic" mediator website, in such a way that each feature exists in its own module. Later, we can then construct other verticals that specialize in each use case.

Current Project State

As of now, the project is only half-complete. Currently, a user can only log in/out, as well as change a few settings. We have a partial implementation of an email notification service, as well as a few UI components for the actual chat itself (though no logic, persistence, or anything remotely AI related). Everything that has been implemented has an extensive testing suite, but there are dozens of race conditions that crop up from time to time, so it can be flaky.

Code is sort of documented, but it might be a bit outdated in spots. There are tons of READMEs explaining usage and architecture, and these will need to be managed and extended as the project goes on. Since this is an ongoing project for students, it's really important that we document all of our design decisions so that future students can carry on after we graduate.

Current Tech Stack

• Typescript + React
  • using shadcn component library
• Vitest (unit testing)
• Playwright (e2e testing)
• Vite
• PNPM (as opposed to NPM)
• Turborepo (monorepo management)
• Supabase (database)
• No current backend
  • planned to run bun (as opposed to node.js)
• Will likely use a ChatGPT model for the AI agents
• not currently deployed

TLDR

• Build a website to help people talk to each other, not at each other
• For a BYU professor and will be used IRL (might even make money later)
• Goal is to deploy the "generic" vertical in a state where its ready to begin testing with real users
• Partially complete, but all the important stuff is left
• Highly modular to eventually build a number of other similar websites
• Project will continue after the class ends (may not have funding to hire anyone though)