What's wicked cool about it?
It's a über-lowlevel approach, the blog post doesn't mention sockets, the author confuses message queueing with SendMessage-style messages, and dismisses memory-mapped files entirely. Also, reading the blog comments, it sounds like the code is buggy and haven't been fixed even though race conditions were mentioned as early as March 15, 2010 - that's almost a year ago. Caveat Emptor.
Only briefly scanned through the code, but it looks like it only handles byte streams and spawns a thread per connection? Ho humm. Not the abstraction level I'd like to work at; thread per connection might be excusable since you're not likely to have a crapload of connections, but for something general-purpose I'd prefer async I/O.
Oh, and it also seems the code requires a connect/data-xfer/disconnect for each "message" rather than supporting connected mode - not good if you need to do high-performance stuff.