Managing a large software project involves initial estimates that may turn out to be erroneous or that might be expressed with some degree of uncertainty. Furthermore, as the project progresses, it often becomes necessary to rework some of the work packages that make up the overall project. Other work packages might have to be abandoned for a variety of reasons. In the presence of these difficulties, optimal allocation of staff to project teams and teams to work packages is far from trivial. This paper shows how genetic algorithms can be combined with a queuing simulation model to address these problems in a robust manner. A tandem genetic algorithm is used to search for the best sequence in which to process work packages and the best allocation of staff to project teams. The simulation model, that computes the project estimated completion date, guides the search. The possible impact of rework, abandonment and erroneous or uncertain initial estimates are characterised by separate error distributions. The paper presents results from the application of these techniques to data obtained from a large scale commercial software maintenance project.