Hindi speakers commonly use a very ambiguous but oddly standardized form of English transliteration which depends on the native speakers ability to guess what the word corresponds to. Here's a random sentence I tried which the tool perfectly translated into Hindi:
Isse aage hum aur kya kahein jaanam samjha karoHere we see that, what Google does yet again, is that it adopts the way people naturally write romanized Hindi, instead of trying to make them learn a new method of transliteration that is easier for computers to parse. This other thing is what Palm did with Graffiti, which was a cumbersome way of writing text, character-by-character, that the PDAs could easily decipher.
इससे आगे हम और क्या कहें जानम समझा करो
If you notice, the Google transliteration engine will not work offline, because it retrieves the transliterated results directly from the Google servers as soon as you hit the space key after typing a word. Although I do not know how the software is implemented, I am guessing it uses large volumes of frequency data as to what English letter combinations correspond to what Hindi letter combinations, sorted by popularity. In that way, they would not need to devise a complicated parsing algorithm for a transliteration system that is standardized only by the weak force of common consensus, but instead present only what seems to be the most probable match for the given word and give the rest of the matches when you click on the word, which, of course, has an uncanny resemblence to "I'm feeling lucky" and "Google Search". It is also the same way that most romanized Chinese and Japanese input systems work.
In fact, what would be really awesome would be if Google converted this amazing piece of web-based software into Input Methods for Mac and Windows so that they could be used in any app. on any computer.