Building a Multi-Model AI Platform with Intelligent Routing

The product problem

A multi-model product must help a user choose how to use several AI models without making every provider detail their problem. It also has to account for usage, present results clearly and connect model work to the account and billing system.

MetaLLM supports direct mode, multi-model querying, intelligent routing and a debate workflow. Those modes place different demands on orchestration, frontend state and cost tracking.

Routing and debate workflows

I built a routing layer that analyzes query complexity and dispatches work to specialized model categories such as reasoning, coding, creative and data analysis. The debate mode runs multiple rounds in which models challenge and refine each other's outputs before a neutral AI judge evaluates the result.

These modes need explicit flow state: which model was called, when a round completed and what the user is paying for. A successful response is more than a single completion from an API.

Application and data architecture

I developed the React and TypeScript frontend using Vite, TanStack Query and shadcn/ui, with an Express.js/Node.js backend. PostgreSQL and Drizzle ORM store the application's relational data. The product includes OIDC authentication and usage-based billing with fiat and cryptocurrency payment paths.

The platform also has a live Android application. The web and mobile clients depend on consistent API behavior and account state across model interactions.

My contribution and outcome

My work covered routing, the multi-round debate feature, full-stack application implementation, database design and usage-based monetization. The product is live; I have not included private subscriber counts, revenue figures or unsupported performance benchmarks.

The public site and repository provide more context about the product and implementation.

Visit MetaLLM ↗ · View source ↗

Related services

AI Agent Development ↗ · SaaS Development ↗ · Full-Stack Development ↗