It might service some of your watched repos because of data quality issues. It uses a very basic nearest neighbor shared item set model. I'm definitely planning on improving the quality of the results in the near future.
Coincidentally, I'm building a github recommendation service as a SNA research project. You may be aware, but for those who are not, GitHub actually had a NetFlix-style recommendation contest in 2009.
Neat in concept, but mostly useless in that it recommended pretty much all projects that I have zero interest in (javascript/coffeescript stuff), which is odd considering I'm not watching any Javascript projects, or at least I don't think I am. It made a couple of shell recommendations, one for IE which I would never touch, one for Spark which I already have installed, and one for Gitflow. Gitflow seems neat though, so I guess one good recommendation might have been worth the rest of the noise?