Libra

Tired of bad arguments? Libra is your fine tuned AI debate mediator, calling out logical fallacies in real-time. Keep debates fair, expose flawed logic, and ensure your arguments are sound.

Comment

Inspiration

Libra began with a familiar frustration: watching debates where smart, confident speakers make claims with no immediate feedback on the strength of their logic or the accuracy of their facts. We set out to build a tireless coach—one that listens to every turn, flags logical fallacies in real time, and fact-checks assertions through live search, all using a single phone passed between two speakers. Our goal is educational, not to serve as an online “truth arbiter,” but to help students develop critical thinking skills and stronger argumentative discipline.

What it Does

At a high level, Libra turns a phone into a turn-based debate mediator. Each speaker gets a timed slot; the app records the audio and uses ElevenLabs for high-accuracy speech parsing before sending it to the backend for transcription, fallacy detection, and fact-checking. Our fallacy pipeline is powered by a fine-tuned AI model built specifically to recognize debate-style reasoning errors. Libra then returns a structured analysis for that turn. After several rounds, it generates AI-written summaries for each speaker, along with a post-debate report that highlights fallacies, inaccurate claims, and overall performance, all stored in our Snowflake-powered database. The experience feels like having a debate coach in your pocket—one that pauses after every turn to say, “Here’s where your logic slipped, and here’s where your facts didn’t hold up.”

How we built it

We built Libra as a full end‑to‑end system

Frontend (React Native + Expo)

  • Single phone, turn-based UI with a recording screen, analysis screen, and summary screen.
  • Audio capture with expo-av, animated timers, and clear state around recording, uploading, analyzing, and ready-to-switch.

Backend (Flask + Python)

REST API endpoints for:

  • /api/transcribe – ElevenLabs speech‑to‑text
  • /api/fallacies – fallacy detection via a custom fine-tuned OpenAI model
  • /api/factcheck – an fact‑checking workflow using OpenAI + web search using an agentic loop
  • /api/generate-summary – GPT‑4‑powered debate summaries per speaker
  • /api/save_debate /
  • /api/get_debate – persisting and retrieving debates via Snowflake

Challenges we ran into

Latency & UX: Chaining STT, fallacy detection, and fact‑checking introduces real latency. We had to: Run fallacy detection and fact‑checking in parallel. Show progressive feedback: transcript → fallacies → facts.

Agent Loop Design: Getting the fact‑checker to: Avoid infinite loops. Know when it has “enough” information.

Accomplishments that we're proud of

Our fact-checker is more than “ask GPT if this is true.” It’s an autonomous agent that:

  • Extracts concrete claims
  • Searches for evidence
  • Evaluates credibility
  • Loops and refines queries

What we learned

  • How to build and demo mobile apps with Expo
  • How to design parallel inference pipelines to reduce perceived latency
  • How to build agentic fact-checkers that balance autonomy and control

What's next for Libra

  • Practice Modes & Personalization
  • Solo practice mode where you debate against an AI opponent that adapts to your level.

Introduce counter‑argument suggestions, where the AI proposes strong rebuttals the opponent could have made.

Share this project:

Updates