NewsBlur is an RSS feed reader that tries to do two things very well:
1) Shows you the original site instead of a context-less feed. Read the original and NewsBlur marks the stories you've read.
2) Filter stories you either like or dislike. A three-stop slider goes between dislike, neutral, and like (red, yellow, and green). Training is super-easy and all click-based (as opposed to you having to writing out what you like in a site, NewsBlur asks you, semi-Hunch-style, your opinions on facets of the site).
I started working on NewsBlur to see if I could do it. Put the AI together with the back-end feed processing and fetching, along with the nifty front-end of the original site. This is one of those projects where I just kept pushing in all directions until I felt I had something good, not knowing if I could do it at all, but believing the entire time that I was able to complete the project.
If you have a project kicking around in the back of your mind, just dedicate enough weeks and months to get it done and done well. Provided you do that, you're almost guaranteed to succeed in some way. Those first few months of writing NewsBlur, I felt like a fraud, wishing I could just get threaded feed fetching to not break every few minutes. Even what I considered to be simple python gave me headaches. (Turns out threading is evil, use multiprocessing instead.)
NewsBlur is entirely open-source: http://github.com/samuelclay/NewsBlur/. The iPhone app, the front-end JavaScript, and the back-end Python. Pick it apart, but know that it is more a learning experience than a business. Turning it into something profitable may not be easy, but I only need 80 premium users to make that happen.