Within machine learning, Peter Cheeseman developed Autoclass, a very popular clustering program which is probably his most highly cited work. Kevin Murphy used to be very well known, when he was a postdoc, for his Bayes net toolbox for Matlab. The author of svmlight is another example; the package was developed during his phd. I can't recall for sure, but I think FFTW started out as a student's project.

The key is to find a modular component that is difficult or novel, and at the same time broadly applicable and hopefully extensible.