Parity

Landing page
Set your api key
Pick side
Choose side
Arena
Request verdict
Verdict

💡 Inspiration

Most AI tools agree with you. They summarize, assist, and validate.

We wanted to build the opposite.

The idea came from a simple frustration: online debate is broken. Social media rewards dunking over reasoning. People never hear the strongest version of the opposing argument — only the weakest. The result is a generation that is very confident and very bad at arguing.

P.A.R.I.T.Y. was born from one question:

What if AI forced you to confront the best possible case against what you believe?

🧠 What It Does

P.A.R.I.T.Y. — Parallel Algorithms for Resolution of Ideological and Tactical Yields — is a full-stack debate intelligence engine.

You enter any proposition. The AI steelmans both sides simultaneously — no straw men, only the strongest possible arguments for each position. You pick a stance. Then you debate across 5 rounds against an AI persona of your choosing. At the end, an AI judge scores the debate and renders a verdict.

🎮 Debate Modes

Mode	Description
Casual	Open sparring, no structure
Oxford	Formal proposition/opposition format
Socratic	AI responds only with questions
Speed	30 seconds per turn
🎙️ Voice Duel	Fully immersive audio arena

🤖 AI Personas

Persona	Style
🏛️ Socrates	Questions everything
⚖️ Attorney	Argues on precedent
🔬 Scientist	Demands evidence
📰 Journalist	Contrarian by nature
📚 Kant	Pure reason only

🏗️ Architecture

System Overview

┌────────────────────────────────────────────┐
│              React Frontend                │
│  TopicInput → Steelman → SideSelect        │
│  → DebateArena / VoiceArena → Verdict      │
└─────────────────┬──────────────────────────┘
                  │ HTTP / SSE streaming
┌─────────────────▼──────────────────────────┐
│           FastAPI Backend                  │
│  /steelman  /debate  /judge                │
│  /fallacy   /coach   /persona              │
│  /n8n       /miro                          │
└──────┬──────────────────────┬──────────────┘
       │                      │
┌──────▼──────┐      ┌────────▼────────┐
│ Gemini 2.5  │      │  Featherless AI │
│ Flash (pri) │      │  (fallback)     │
└─────────────┘      └─────────────────┘

Agent Pipeline

User Input
    │
    ├──► debate_agent     → streams AI response (SSE)
    ├──► fallacy_agent    → detects logical errors in parallel
    └──► coach_agent      → computes coaching hints (on demand)

After final round:
    └──► judge_agent      → scores and renders verdict
         └──► n8n_agent / miro_agent  (on user request)

📊 Scoring Model

The judge evaluates three dimensions independently, then computes an overall score using a weighted harmonic mean — to penalize any single weak dimension:

$$S_{overall} = \frac{3}{\dfrac{1}{w_L \cdot S_L} + \dfrac{1}{w_E \cdot S_E} + \dfrac{1}{w_C \cdot S_C}}$$

Where:

$ S_L $ = Logic score $\in [0, 100]$
$ S_E $ = Evidence score $\in [0, 100]$
$ S_C $ = Clarity score $\in [0, 100]$
$ w_L = 0.40,\; w_E = 0.35,\; w_C = 0.25 $

Winner is determined by:

$$\text{Winner} = \begin{cases} \textbf{User} & \text{if } S^{user}{overall} - S^{ai}{overall} > \delta \ \textbf{AI} & \text{if } S^{ai}{overall} - S^{user}{overall} > \delta \ \textbf{Draw} & \text{otherwise} \end{cases}$$

Where $ \delta = 5 $ is the minimum margin for a decisive victory.

🗡️ Steelman Strength Ranking

Each argument receives a strength score from three normalized sub-scores:

$$\text{strength}_i = \alpha \cdot P_i + \beta \cdot R_i + \gamma \cdot N_i$$

$ P_i $ = Persuasiveness
$ R_i $ = Relevance to the proposition
$ N_i $ = Novelty (penalizes redundancy with prior arguments)
$ \alpha = 0.5,\; \beta = 0.3,\; \gamma = 0.2 $

⚠️ Fallacy Detection

Every message is scanned in parallel. A fallacy is flagged when:

$$\max_{j \in {1..k}} P(\text{fallacy}_j \mid \text{argument}) > \tau$$

Where $ \tau = 0.72 $ — tuned to minimize false positives while catching genuine logical errors.

Detected fallacy types include: Ad Hominem · Straw Man · False Equivalence · Hasty Generalization · Slippery Slope · Appeal to Authority

🎙️ Voice Duel — State Machine

The Voice Arena runs as a finite state machine with four states:

$$Q = {\textbf{Idle},\; \textbf{Listening},\; \textbf{Thinking},\; \textbf{Speaking}}$$

$$\delta(\text{Idle}, \text{tap}) \rightarrow \text{Listening}$$ $$\delta(\text{Listening}, \text{silence}_{t > 1.2s}) \rightarrow \text{Thinking}$$ $$\delta(\text{Thinking}, \text{response_ready}) \rightarrow \text{Speaking}$$ $$\delta(\text{Speaking}, \text{tap} \cup \text{utterance_end}) \rightarrow \text{Idle}$$

Interrupt logic calls window.speechSynthesis.cancel() on any tap during Speaking — preventing the AI from repeating cancelled phrases.

🛠️ How We Built It

Backend

Python + FastAPI with 8 specialized agents. All AI calls route through a unified client using the OpenAI-compatible API — allowing zero-code provider switching between Gemini 2.5 Flash (primary) and Featherless AI (fallback) via a single base_url swap and a ContextVar for per-request key isolation.

Frontend

React + TypeScript + Vite. Framer Motion for all transitions. Space Mono + Space Grotesk for the terminal-meets-modern aesthetic. Every screen is a purpose-built component with its own Zustand state slice.

🎙️ Voice Duel Arena

A dedicated full-screen immersive audio component — entirely separate from text chat. Uses the Web Speech API for both recognition and synthesis, with persona-matched voices: Socrates → British English, Kant → Irish, Journalist → American.

Sponsor Integrations

Integration	What It Does
⚡ n8n	One-click exports full debate + scores to a webhook → email, Discord, Notion, Sheets
🗺️ Miro	Generates a live visual argument map — every claim, counter, and fallacy on a board
🪶 Featherless AI	OpenAI-compatible fallback inference — activates automatically when primary limits hit

Deployment

Google Cloud Run via Google Cloud Buildpacks. API keys injected as environment variables at deploy time — never in code or version control.

⚔️ Challenges

1. Cloud Build — Eight Failures Before Launch

Unicode box-drawing characters (═══, ─) in the Dockerfile corrupted Cloud Build's parser on Linux
Google Buildpacks selected Python 3.14 (bleeding edge), breaking dependency resolution
Concurrent IAM SetIamPolicy writes caused ABORTED (gRPC code 10) race conditions
Billing not activated → CreateBuild silently returned FAILED_PRECONDITION (code 9) with no useful error message in the CLI

2. Voice Duel State Integrity

Browser SpeechRecognition.onend fires unpredictably — sometimes mid-sentence. Getting the FSM right required careful debounce tuning and explicit interrupt guards to prevent feedback loops.

3. Fallacy Detection Calibration

Threshold $ \tau $ required iteration across 200 test arguments. Too low → every argument flagged, destroying trust. Too high → genuine errors slip through. Final value: 0.72.

4. Steelman Quality

Preventing shallow or repetitive arguments required explicit novelty constraints in the prompt — each subsequent argument must introduce a new dimension not covered by prior arguments in the same panel.

🏆 What We Learned

The steelman constraint is transformative. When you force the AI to present the strongest possible opposing case — not just any rebuttal — the quality of the entire debate goes up. You cannot dismiss it. You have to actually engage.

Debate is a solvable UX problem. The reason online discourse is so poor isn't that people are irrational — it's that the tools reward speed and outrage over precision. Give someone a structured 5-round format, a live fallacy detector, and a scored judge, and reasoning quality improves measurably.

And finally: never put Unicode characters in a Dockerfile.

Built for LovHack Season 2 · Powered by Featherless AI × n8n × Miro