'Pipeline running' delay is a pain, something potential users/customers won't wait around for. Have you looked at using on-demand managed containers (like Google Cloud Run) instead of full EC2s to launch the pipeline? It'll be cheaper, quicker and more scalable too.

Just saw that JSON support already exists, which is great.

Also, it has the potential of being a great platform, but I wouldn't suggest marketing it as a 'Roblox for AI'. How many AI/ML people really use Roblox? Few I suppose. Perhaps something simpler and familiar, like 'GUI for AI' or 'Your ML Lab/Studio/Toolbox'.