Ah yes, I dived right in there 
The .dist and Potential( logp() ) is a fairly common alternative pymc3 idiom to replace ‘observed=’. I pasted it here from a model I had to hand - in which I found that observed= led to errors. You can likely safely ignore this idiom for now and use the observed= pattern e.g. Multivariate — PyMC 5.10.0 documentation
More generally in my example, I have a different MvNormal for each of nobs, and where they share a common, prior, Cholesky-decomposed covariance matrix.
If you wanted to pick a normally distributed value of alpha and also of beta, per observation, then this structure would give that. And the LKJCholskyCov prior would ensure that these values covary.