Consensus

A Debate Orchestration System that offsets your worries and helps you reach clarity amidst your toughest dilemmas. Start a debate and enjoy an audio-first experience of agentic chaos AND resolution.

Comment

Inspiration

The inspiration for Consensus came from the realization that decision-making often benefits from diverse perspectives. However, accessing such perspectives in a structured, meaningful way is challenging. So I created a tool that could simulate this process, offering users clarity through orchestrated disagreement. Imagine convening a panel of diverse experts to debate your toughest dilemmas. Type in a juicy question (or use one of the sample prompts!) and witness all personas pull in opposite directions just to synthesize an insight together with the help of a moderator.

What it does

  • Accepts a decision or dilemma from the user
  • Collects a small amount of context to frame the debate
  • Runs five personas in a structured debate, then asks a moderator to synthesize the result
  • Plays persona responses as audio using ElevenLabs
  • Includes a follow-up flow so users can ask the moderator for more guidance.

How we built it

Frontend: Vanilla HTML, CSS, and JavaScript for a lightweight, responsive interface.

AI: Google Gemini 2.5 Flash for generating persona arguments.

Audio: ElevenLabs Text-to-Speech for high-quality voice synthesis, with each persona having its own characteristic voice.

Backend: Node & Express for setting up a server and safely routing the API Keys.

The debate flow is engineered to run persona responses in parallel, minimizing wait times. Audio synthesis is triggered as soon as responses arrive, ensuring playback is ready when the user clicks play. The system is designed for extensibility, allowing for easy addition of new personas or features.

Challenges we ran into

  • Ensuring the personas felt distinct and authentic. This required fine-tuning their system prompts and audio settings.
  • Integrating multiple APIs (Gemini, ElevenLabs) while maintaining a smooth user experience.
  • Race conditions, ensuring audio was ready at the right moment, balancing tradeoffs.
  • Accidentally uploading an API key to GitHub; learned how to remove it from commit history

Accomplishments that we're proud of

  • First hackathon; proud of making even a valid submission given how things were collapsing at one point.
  • Building a fully functional system that combines AI and audio technologies. Successfully integrating Google Gemini for persona-driven arguments and ElevenLabs for high-quality voice synthesis was a significant milestone.

What we learned

Building Consensus taught me the importance of orchestrating multi-agent systems effectively. I learned how to fine-tune AI personas to ensure diverse and meaningful perspectives while maintaining coherence. Integrating multiple APIs required overcoming latency and compatibility issues.

What's next for Consensus

The next step for Consensus is to expand its capabilities meaningfully. Before introducing more personas, I'd focus on integrating memory to keep track of past reflections. With that would come authentication and security enhancements. Or hijack it as a personal learning project to Frankenstein upon with new tools that interest me.

Share this project:

Updates