Thank you for the article. I've been a long time proponent of JSON. I've implemented something similiar in the CLR and I have a couple suggestions for handling datasets over 100 rows to get the performance up, size down, memory utilization down and retain complex data types:

1) Put a Data map at the start of your structure that stores the names of the columns size and type. Cast and validate on the client side if necessary.

2) Remove all the Key names from every entry and utilize the map and utilize standard nested arrays

Thank you,

Blaine