After doing some timing measurements, I believe using Nilsimsa in a bulk mail detection scheme would be practical on a machine with lots of memory (about half a gigabyte should do it).
The problem with this idea is that Nilsimsa's statistical properties when fed English text (or text in any other language, for that matter) aren't the Gaussian approximation to a binomial which you'd really like them to be able to work out useful measurements of how two texts differ. This effectively kills it as a fuzzy checksum, alas. This information is left here purely for historical reasons: I'm no longer looking at Nilsimsa myself.