Workload routing control for intent-native infrastructure.

A compact routing device for cloud and on-prem compute. It reads workload intent, identifies the optimal substrate (CPU, GPU, TPU, FPGA, serverless, edge, batch, distributed), and hands execution to the right infrastructure.

engine

<1ms

footprint

26MB

routes

mode

compare-only

Live Router

state the workload

Natural language in. Substrate out.

Use ARBITER as the pre-routing layer above cloud APIs and infra schedulers.

intent

standby

selected substrate

—

coherence

—

gap

—

latency

—

Quickstart

integration

Install the routing layer.

ARBITER does not replace schedulers. It converts user language into a substrate class, then your application calls cloud APIs, Kubernetes, or batch systems.

intent → class → schedule → execute

Collect intent Workload type, latency, parallelism, determinism, cost constraints.

Call ARBITER Send intent + compute substrate candidates to /compare.

Map substrate class CPU, GPU, TPU, FPGA, serverless, edge, batch, distributed, placement.

Schedule and execute Use selected substrate to provision resources and run workload.

request/public/compare

Compute Catalog

Compute substrate directory.

CPU, GPU, TPU, FPGA, serverless, edge, virtual machines, bare metal, batch, distributed (Spark, Dask, MPI, Ray), placement (spot, reserved, carbon-aware), and orchestration.