The machine learing service provides a common runtime for evaluating machine learning models on device. The service wraps the TensorFlow runtime which has been optimized to support the set of built-in machine learning models with are installed on rootfs. See go/chromeos-ml-service for a design overview.
The Mojo connections are bootstrapped by passing the file descriptor for the IPC pipe over D-Bus. The life-cycles for the loaded machine learning models and the runtime eval session are matched to the life-cycles of the Mojo interfaces. See go/chromeos-ml-service-mojo for a description of the Mojo interfaces and go/chromeos-ml-service-impl for details on bootstrapping the connection over D-Bus.