In your case, if you do posterior prediction using pm.sample_posterior_predictive, you can get the predicted observation, and you can use these posterior predictive samples to estimate the uncertainty (computing the variance, for example). Is it what you are looking for?