What inspired it? I kept building the same thing over and over. At my day job, I was standing up AI research tools for a large organization and every new project started with the same slog: wire up document storage, chunk everything, embed it, build a retrieval layer, add hybrid search, expose it somehow. None of that was the interesting part. It was just infrastructure tax. The problem. The bigger frustration was that most document backends treat a PDF or a research report as a blob of text. But those documents have structure — sections, abstracts, methodologies, conclusions. When a user asks a question, the right answer often lives in a specific section, not just the nearest chunk. Existing tools either ignored that or made you build it yourself. How the approach evolved. I started with the RAG problem and kept pulling the thread. If I was going to build a proper document backend, it needed real-time sync (so UIs stay live), hybrid search as the default (not an afterthought), and a way for AI agents to talk to it natively. That last piece led me to MCP, and the idea of a dedicated MCP server per collection clicked as the right abstraction. Suddenly agents could treat a document collection like a first-class data source without any custom glue code.

Dewey

The real-time document backend built for AI apps

The real-time document backend built for AI apps