Hi HN. I built AskVerdict after a week at Anthropic's Claude Code Hackathon.
The core idea: instead of one LLM giving you one answer, spawn multiple agents with distinct stances that debate the question through structured rounds. Forced cross-examination, argument status tracking (standing/challenged/rebutted/conceded), and a synthesized verdict with confidence scores.
Under the hood it's a TypeScript monorepo: Hono + Bun for the API, Next.js frontend, Postgres with Drizzle ORM. The debate engine uses a multi-tier model router — heavier models for orchestration and synthesis, lighter ones for debate rounds and classification. Keeps costs under $1 per debate.
BYOK only right now. Bring your own API keys for Claude, GPT, DeepSeek, Groq, or OpenRouter. No vendor lock-in.
Opening to the first 100 beta users. Feedback welcome, especially on the debate protocol design — should agents be allowed to concede gracefully or should the system force them to defend until the end?
Hi HN. I built AskVerdict after a week at Anthropic's Claude Code Hackathon.
The core idea: instead of one LLM giving you one answer, spawn multiple agents with distinct stances that debate the question through structured rounds. Forced cross-examination, argument status tracking (standing/challenged/rebutted/conceded), and a synthesized verdict with confidence scores.
Under the hood it's a TypeScript monorepo: Hono + Bun for the API, Next.js frontend, Postgres with Drizzle ORM. The debate engine uses a multi-tier model router — heavier models for orchestration and synthesis, lighter ones for debate rounds and classification. Keeps costs under $1 per debate.
BYOK only right now. Bring your own API keys for Claude, GPT, DeepSeek, Groq, or OpenRouter. No vendor lock-in.
Opening to the first 100 beta users. Feedback welcome, especially on the debate protocol design — should agents be allowed to concede gracefully or should the system force them to defend until the end?