Something that I found very interesting and very useful is [Singular value decomposition][1].
It shows that every operator is "almost diagnosable", and is skipped in a lot of basic linear algebra courses I have seen.

I has many application, for example - solving sum of least squares of example. You can give a 30 minute talk on this in various levels as well.

There are prettier theorems (Stokes, Uniformization, and many more) but I think with the 3 constraints (interesting, useful, little background) this is a good topic.  


  [1]: http://en.wikipedia.org/wiki/Singular_value_decomposition