Skip to content

Adding new models to the platform

This section covers the full process of integrating a new analysis model into ELSI, from implementation to deployment. The process involves two actors: the model developer who prepares and submits the model, and the platform maintainers who deploy it.

Model Developer: Implementation

Dockerizing the Model

All models must be packaged as Docker images to run on the platform. The implementation must follow the requirements defined in the analysis-service-core repository, which specifies the interface your model must expose. This includes:

  • GPU support — implement if the model can benefit from it
  • Failure reporting — the model must report errors in the expected format so the platform can surface them to users
  • Effort model — an estimate of the compute cost relative to input duration, used to calculate credit consumption
  • Progress reporting — the model must report progress so users can track long-running jobs

Read the repository documentation carefully before starting — the interface requirements are strict and must be met for the model to be accepted onto the platform.

Annotation Output Format

Check whether your model's output format is already supported by ChildProject by consulting the list of supported converters.

If your format is not on that list, you must add a conversion step inside your Docker image that transforms the model output into the standard ChildProject annotation CSV format. The required fields are documented in the ChildProject annotation format reference.

Submitting to the Platform

Once your Docker image is built and tested, submit it to the platform maintainers. Provide:

  • The Docker image or instructions to build/pull it
  • The name to use for the service in the Swarm compose file
  • A short description of the model and its expected output format