What we build
Our work ships in Tellura, our flagship product. It brings together two systems: a co-author for people writing original fiction, and a pipeline for translating long-form work.
The co-author
A general-purpose writing assistant for authors. It plans, drafts, and revises alongside a human writer, keeping track of a story's world and characters as the work grows. It is built to help people create original fiction. It is not a translation tool.
The translation pipeline
A system for translating long-form fiction across languages. Rather than handing a chapter to a language model and hoping for the best, it runs translation through a deterministic, verifiable process with explicit rule-based checks: holding terminology, facts, and a story's world consistent from the first chapter to the last. We describe it as neuro-symbolic: language models do the creative work, and deterministic rules keep them honest.
The translation pipeline is built on an architecture we call a Recursive Language Model: a way of orchestrating language models through a deterministic process, not a model we train ourselves.
The idea behind both
These two systems are deliberately different. The co-author is open-ended: it is meant to explore, suggest, and follow a writer wherever the story goes. The translation pipeline is deliberately constrained: its job is fidelity, so it is held to explicit rules and checks. The distinction is where we let a model drive, and where we refuse to.