Hi, and welcome 
It seems to me that you’re turning to a Bayesian model because you care about the uncertainties in your estimates. If yes, then my question would be: why aren’t you running MCMC to get full posterior distributions of your parameters, instead of using find_MAP, which will only give you point estimates?
Also, note that initalizing the MCMC sampler at the MAP is discouraged now – just let PyMC initialize the sampler for you.
Hope this helps 