Startup Jury AI — Project Documentation

Overview

Startup Jury AI (codename: Argument Arena) is an interactive web application that simulates a startup pitch panel. A founder describes their startup idea, and a panel of 8 AI-powered expert personas debate the idea across up to 4 rounds. After each round the founder can respond to the panel’s questions and challenges. At the end, each panelist scores the idea using their own weighted criteria, and a Consensus Judge delivers a final GO / MAYBE / NO-GO verdict.

Tech Stack

Application Flow

Setup → Debating (rounds 1–3) → Final Ratings (round 4) → Judge Verdict  

Phases

Debate Rounds

• Round 1 — Each panelist reacts to the raw idea (≤ 100 words each).
• Rounds 2–3 — Panelists react to each other and the founder’s response (≤ 75 / 50 words).
• Round 4 (Final Ratings) — Panelists give a verdict and a scored breakdown using their individual weighted rubrics.
• Judge — A neutral Consensus Judge aggregates all scores into a GO / MAYBE / NO-GO verdict.

After each round (except the last), the founder can type or speak a response to the panel before triggering the next round. An Auto-Debate mode can automatically generate a simulated founder response, allowing the full debate to run without any user input.

AI Personas

All 8 personas are defined in src/data/personas.ts. Each has a unique character, scoring rubric, and system prompt.

Inverse scoring: Riley Novak (Skeptic) scores the risk level of an idea. A raw score of 10 means extremely risky; the system inverts it (10 - score) for the final tally.

Verdict Thresholds

Key Source Files

Key Components

Getting Started

Prerequisites

• Node.js ≥ 18
• A Supabase project with:
  • debate-ai Edge Function deployed
  • copilotkit Edge Function deployed
  • Environment variables set

Environment Variables

VITE_SUPABASE_URL=https://your-project.supabase.co  
VITE_SUPABASE_ANON_KEY=your-anon-key  

Install and Run

npm install  
npm run dev  

Other Scripts

npm run build      # Production build  
npm run preview    # Preview production build  
npm run lint       # ESLint  
npm test           # Run tests (Vitest)  
npm run test:watch # Watch mode tests  

AI Architecture

All AI calls flow through src/lib/ai.ts via the callCompletion helper, which invokes the debate-ai Supabase Edge Function. That function forwards prompts to a Claude model and returns the completion.

Prompt Strategy

• Each persona has a distinct systemPrompt defining their character, priorities, and scoring rubric.
• Industry keywords in the idea (e.g., “fintech”, “health”, “saas”) are auto-detected and injected into prompts.
• Word limits shrink each round (100 → 75 → 50 words) to keep debates focused.
• Memory: each persona accumulates notes across rounds via Redis, enabling continuity.

CopilotKit Integration

The debate state is published to CopilotKit via useDebateAgentState, allowing a CopilotKit agent to observe the live debate and emit AgUI events (e.g., user_response). The runtime endpoint is the copilotkit Supabase Edge Function.