Delibero

Delibero is an AI-powered strategic debate platform where multiple expert personas deliberate your toughest business questions. It uses adversarial reasoning to challenge assumptions, stress-test ideas, and synthesize a more robust path forward.

🚀 Features

• Strategic Reframing: Before the debate begins, a specialized agent analyzes your question to identify hidden asymmetries and reframe the problem statement.
• Multi-Persona Debate: Four distinct AI agents (CEO, CFO, CTO, Product) debate the reframed question, each with their own incentives and cognitive biases.
• Adversarial Reasoning: Agents don't just agree; they are prompted to challenge each other, expose risks, and fight for resources.
• Constraint Critic: A "reality check" phase that evaluates the emerging consensus against your specific real-world constraints (budget, timeline, team size).
• Strategic Synthesis: The final output isn't just a summary—it provides a clear Strategic Thesis, a decision matrix, and Lessons from the Losing Side.

🧠 The Personas

1. Elena Vasquez (Chief Strategy Officer): Visionary, focuses on long-term moats and market positioning.
2. Morgan Blackwell (CFO): Conservative, focuses on capital preservation, unit economics, and option value.
3. Dr. Marcus Chen (Devil's Advocate): Contrarian, specifically targets cognitive biases (sunk cost, action bias) and organizational incentives.
4. Sarah Jenkins (VP Strategy): Pragmatic, focuses on execution feasibility and customer impact.

🛠️ Tech Stack

• Framework: Next.js 15 (App Router)
• Styling: Vanilla CSS (CSS Variables, Flexbox/Grid) — no Tailwind
• AI: OpenAI (GPT-4o) or Anthropic (Claude 3.5 Sonnet)
• Streaming: Server-Sent Events (SSE) for real-time debate playback

🏗️ Technical Architecture

1. Hybrid Orchestration (Client-Driven)

Delibero uses a unique "Thick Client, Smart Edge" architecture to handle complex, long-running debates without timeouts.

• Orchestrator: The debate loop (engine.ts) runs partially on the client to manage state and partially on the Edge to generate tokens.
• Streaming: All agent responses are streamed in real-time using ReadableStream, ensuring the application remains responsive even during 60+ second generations.
• Persistence: Debate history is stored entirely in the user's browser (localStorage), prioritizing privacy and removing the need for a database.

2. The Reasoning Topology

The core engine follows a strict, multi-stage topology for every round:

1. Generation: Agents generate arguments in parallel based on their persona constraints.
2. Evidence Audit: A separate model extracts claims and checks for hallucinations or contradictions.
3. Cross-Examination: If contradictions are found, a "Cross-Exam" phase is triggered where agents interrogate each other.
4. Constraint Critic: An iterative loop checks the emerging consensus against hard constraints (Budget, Legal, Timeline) before the next round starts.
5. Synthesis: A final step compresses the round's context and decides if the debate needs to diverge (extend) or converge (end).

3. Edge-First API

• Runtime: /api/chat runs on Vercel Edge Runtime to bypass the standard 10s serverless timeout, allowing for deep-thinking models to run to completion.
• Security: Stateless API design protected by a DELIBERO_ACCESS_CODE to prevent unauthorized usage.