Machine learning and AI require lots of storage (for storing data, model checkpoints, production models, etc.) and lots of compute (for experimenting with and training different models, serving the models in production, etc.). Filebase and Akash respectively provides both at a fraction of the cost of their centralized counterparts, but there is no AI/ML platform to take advantage of them right now. Let’s change that.
For this hackathon, I propose to build the beginnings of a DeCloud AI Platform, similar to offerings in tradition cloud like Google AI Platform, Amazon SageMaker, Azure ML, etc. Given the limited time of the hackathon, I will be focusing on implementing a web interface that enables what I think are some of the core features common to most traditional offerings:
- From the web interface UI, users will be able to specify and deploy a custom compute instance that comes with a hosted Jupyter notebook environment at the click of a button. The ability to specify GPUs will come in the future when Akash supports it.
- Within their Jupyter environment, users will be able to load datasets from Filebase S3 and train their models. Model checkpoints and final models will also be uploaded to Filebase. Users can close their Jupyter deployment and the models will still be available from Filebase.
- From the web interface, users will be able to see all their models saved to Filebase. There will be a UI that allows users to pick a model to deploy as an endpoint, so that consuming ML apps can use the endpoint to perform model inference.
Such a website will interface with the Keplr wallet extension for Akash deployment payments, and will be powered by AkashJS.
