Basically it receives the tweets from stream and using Lucene's NRT, indexes the tokens from tweets using N-Grams. By searching for a hashtag, it receives the top 5 keywords describing the hashtag at that moment of the indexing.
Hashtag: #VoteLeave
Context: [leave, vote]
It might not scale in the long run. Most of the index is memory based. I couldn't test it out yet.