How about training a neural network for each user? You could provide lots of inputs that wouldn't be useful for all users (time of day, pervious likes, week of month, moon cycle, volume on twitter, DJIA movement) but may be very useful for a given user...
So many web apps (due to their nature, I'm sure) are based on a one-size-fits-all approach recommendation engine (i.e. Netflix, Pandora). How about an "everybody is different" approach that takes advantage of cloud processing/storage and scale? Sounds like a fun experiment, at least.
Many companies conflate having an API with being a "true platform."
Additionally, we are all witnessing the pussification of the word platform. It used describe something that was more fundamental, more plumbing and infrastructure that others used to build separate and often wholly distinct businesses (i.e. Farmville on the Facebook Platform). But the new watered down platform is any web app with an API that allows mixing Nyan Cats with Foursquare and raising $42MM.
Most web apps should provide an API. Almost none of those will become platforms. So be it.