
Majlis is a multi-model "discussion arena" designed to eliminate AI blind spots and hallucinations. Instead of relying on a single LLM, Majlis runs several open-source models—such as Llama 3.1, Qwen 2.5, and DeepSeek R1—simultaneously on the same prompt. A specialized Curator AI acts as an intelligent router, recommending models based on live reputation scores and flagging shallow or repetitive responses to ensure high-quality, accountable outputs. Technically, the platform leverages the AMD Developer Cloud, running four vLLM inference servers on an AMD MI300X GPU via ROCm. The backend is built with FastAPI and Docker Compose, while the frontend utilizes React 19 and Tailwind CSS v4 for a responsive, real-time moderation interface. By allowing users to dismiss unhelpful models and update their category-specific scores, Majlis creates a self-optimizing ecosystem where the best AI minds earn their seat at the table. Stack Summary: Inference: vLLM on ROCm (AMD MI300X) Models: Llama 3.1, Qwen 2.5, Mistral 7B, DeepSeek R1 Backend: FastAPI, SQLModel, Tavily (Search) Frontend: React 19, Vite, Clerk (Auth) Infrastructure: Cloudflare Pages & AMD Developer Cloud
10 May 2026