I am not able to mentor, but I have some ideas about easier projects. These may be too easy, too hard, or not even desirable so take them or leave them as you please.
Implement a set of circular statistics functions comparable to those in R or MATLAB circular statistics toolbox.
Either implement some window functions that only apply to the beginning and end of an array, or implement a wrapper that takes a window function and some parameters and creates a new window that only applies to the beginning and end of an array.
Integrate the bottleneck project optimizations into numpy proper.
Integrate as much as possible the matplotlib.mlab functionality into numpy (and, optionally, also scipy).
In many places different approaches to the same task have substantially different performance (such as indexing vs. take) and check for one approach being substantially slower. If it is, fix the performance problem if possible (perhaps by using the same implementation), and if not document the difference.
Modify ufuncs so their documentation appears in help() in addition to numpy.info().