Hadoop at Twitter

I have a project, http://elev.at - it's a web API to convert data from HTML tables, Excel spreadsheets, CSV files, etc. into XML - the premise is to make information that is published for humans eyes, into computable form so that other apps can consume the data.

So far I'm only supporting XML as the output, but am wondering about other outputs. One obvious format would be JSON, but I'm also wondering about others, such as Avro. It is similar in purpose as Thrift or Protocol Buffers, but unlike these two, you don't need to pre-gen the code files to process the resulting data sets, i.e. the generation and consumption of Avro data sets can be more dynamic than with Thrift or Protocol Buffers.

I was wondering whether anybody here has experience with Avro.