In the simplest cases of disputed paternity we will have evidence in the form of DNA profiles from the child, mother and putative father: it is then relatively straightforward to determine the strength of the evidence bearing on paternity. But often the putative father is unavailable for testing, and instead we have to make do with DNA from one or more of his relatives. Other features, such as mutation, silent alleles, laboratory and handling errors, etc., introduce additional complications. The task of interpreting the forensic evidence in such cases can become extremely challenging, both logically and computationally. Recently it has been shown how the technology of Bayesian networks -- especially in its "object-oriented" version -- can be used to represent and solve such problems. I will describe the basics of this approach, present a collection of fundamental networks that can be flexibly combined (like a child's construction kit) to represent a very wide range of problems arising in forensic genetics, and illustrate their use in some real cases.