Stanford published a paper that is basically the exact same model: http://crypto.stanford.edu/PwdHash/pwdhash.pdf

This is not a new technique. In addition to the Stanford paper, there are several other implementations mentioned in these comments. It's a compromise, not a mistake. It is better to memorize one strong password than a dozen weak ones.

This isn't custom crypto. It's a well-known hash function that serves as a filter, transforming the passwords you would otherwise enter directly into a website's login form. It is no less secure than typing in passwords by hand.