The geospatial world is full of such datasets where its hard to know exactly how the input variables to your model will effect the outcomes. There exists a growing ecosystem of libraries and frameworks like Tensor Flow and Scikit-Learn that allow for sophisticated machine learning to take place but very few are easily interoperable with geospatial frameworks like PostgreSQL.. In this talk I will discuss ongoing work at CartoDB to integrate machine learning as a key analysis tool for geospatial data. Focusing on our work using random forests, neural networks and Markov chains I will talk about how these methods need to be adapted to work with geospatial data, how we can use the PL/Python extension in PostgreSQL to bring the power of these models to our geospatial data sets and discuss kinds of new analysis these methods open up In particular I will discuss about our work to develop segmentation models that are able to take a set of example observations and train a predictive model based underlying multivariate geospatial datasets like the census and use this model to predict new observations in regions where the original data was missing..