A platform for orchestrated, multimodal AI.
SutraMesh is structured as a small set of cooperating layers. Each layer is an open research problem we are studying and prototyping in isolation before composing them together.
Six layers, one mesh.
Router Layer
A lightweight intelligent router classifies intent, decomposes a request into subtasks, and assigns each one to the most appropriate expert model.
Expert Model Layer
A registry of specialized models — language, vision, code, structured-data, retrieval — each tuned for a narrow domain and addressable as a capability.
Multimodal Coordination
A shared context layer that lets text, image, audio, and document subtasks reason over a common representation as the pipeline executes.
Distributed Inference
Independent subtasks run in parallel; dependent ones run sequentially. The mesh schedules execution and merges partial results into one response.
Workflow Engine
Multi-step, long-horizon agents that plan, act, and self-correct — built on top of the router and expert layer as a higher-order primitive.
Runtime & Observability
Caching, tracing, policy controls, and routing telemetry — the substrate that makes orchestrated AI systems debuggable and reproducible.
What we optimize for.
We are not chasing the largest model. We are studying the architecture around models: how requests flow, where specialization helps, when parallel execution is safe, and how to keep the whole system interpretable as it grows.
Three principles guide our design work today:
- Composition over scale. A coordinated set of smaller models can outperform a single giant model on real workloads — and cost a fraction to run.
- Routing is a first-class problem. Most production AI systems treat routing as glue code. We treat it as the central research question.
- Multimodal by default. The mesh should reason over text, image, audio, and structured data inside one shared pipeline — not bolt modalities on as separate products.
Status of each layer.
See the architecture in detail.
Diagrams of the routing core, expert communication, dynamic task decomposition, and parallel execution model.