A research project on how AI systems should be built.

We started SutraMesh because we believe the most interesting open problem in AI is not how to train a single, larger model — it is how to make many specialized models reason together as one system.

Today the field is mostly organized around the first question. Bigger models, longer contexts, more training data. The second question — how those models should be orchestrated — is still mostly hand-written glue.

We think orchestration is the more important research direction for the next decade of AI infrastructure. The mesh we are building is our attempt to take it seriously.

A lightweight intelligent router that decomposes a multimodal request into typed subtasks, dispatches each one to the specialized expert model best suited for it, executes independent branches in parallel, and aggregates the results into a single coherent response — with the full execution trace exposed as a first-class artifact.

We are not running a product. We are in active research, architecture exploration, and foundational infrastructure development. The website you are reading is the public surface of that work.

We share progress in three places: the research areas page, the roadmap, and the blog when posts are ready.

For research conversations, collaborations, or feedback, email hello@sutramesh.in.

We are a small team. Replies may take time, but every thoughtful message gets read.