Architecture🔗

Kilroy is divided into multiple parts, each with their own responsibilities. You can replace any of these parts with your own implementation, as long as it conforms to the API.

Architecture diagram

Parts🔗

You can read more about each part in the following sections.

Face🔗

The face is the part that is responsible for communicating with a social media platform.

Think of it as a wrapper around the social media platform's API, that conforms to the kilroy API.

There are default implementation for Discord and Twitter. You should write your own face if you want to use kilroy with a platform that is not supported by default.

Module🔗

A module is a part responsible for handling all operations related to the model, like generating output and updating weights.

Again, think of it as a wrapper around the model of your choice that conforms to the kilroy API.

There is a default implementation for any models that are supported by HuggingFace. You should write your own module if you want to use kilroy with your own model.

Controller🔗

The controller is the part that controls the training process using the face and the module. It also serves as a gateway to the system from outside (e.g. from the Web UI).

The default implementation can be found here and should be sufficient for most use cases. But of course, you can write your own controller.

gRPC-Web Proxy🔗

This parts enables the dashboard to communicate with the controller. You can find the implementation here. It should work as it is.

Dashboard🔗

The dashboard is a web app that can be used to control the training process. Think about it as the easiest to use interface for the whole system. The default implementation can be found here.

Communication🔗

The communication between the parts is done using gRPC. The gRPC definitions can be found here. You can use them to generate stubs that conform to the API for your own implementations in any language.