Just thinking out loud here…
I wonder if you could save cached computations away onto disk so the client can use the same already completed computations/model as the researcher (like training a model)
Otherwise, the researcher needs a streamlit app to perform modeling computations Jupyter notebook style and then needs to take their model and create a client application to demonstrate the model on less powerful machines that other stakeholders may be using.
Maybe application states could be used to allow the researcher to work in the app in a development capacity where the app locally runs heavy computation while writing models to the disk/git directory so that the client mode of the app can be run on a less powerful computer or on a low power web server. After all, getting a result from a trained model is 1000x+ less computationally intense than training the model.
You could always rebuild the model from the development tab/screen, but otherwise, the model is already prepared, reducing latency for those who want to toy with a model.