As a general rule, campaigns and state parties outsource the creation of partisanship models to third party vendors like Civis. A partisanship model predicts the likelihood that any given voter affiliates as a Democrat; such models are very useful for prioritizing outreach and canvassing efforts.
Hudson Cavanagh built a partisanship model for the Texas Democratic Party that was maintained in-house. He describes this model in a fair amount of detail in his Great Battlefied interview on partisanship models - at about the 10:08 mark.
One of the most compelling reasons to house partisanship models within state parties is that state parties have unique access to all of the canvassing data in their instance of VoteBuilder. For example, the Texas Democratic Party has access to all of the canvassing data in VoteBuilder Texas database. As a result, they can rapidly iterate the partisanship model, taking the latest field data into account in each iteration of the model; that can reduce the lag time and cost in refreshing a model substantially. Moreover, state parties are legally permitted to coordinate with a wide variety of campaigns within a state - as a result, their work can be readily shared up and down the ballot.