This is good example of how to answer 'big data' questions without big expensive distributed systems via the magic of lots of RAM and probabilistic data structures.
Are you using this in Production now? I took a stab at something similar in Ruby (https://github.com/doomspork/Orwell) awhile ago, seeing this makes me want to dust it off and add new features. Thanks for sharing!
One of the first things I noticed when I checked it out. Both MixPanel and Codecademy are YC so I wonder whether there was any communication around it.