Forgive me if this is a dumb idea, but does the state have any published schema for this document -- can you just get the XSD and use that to generate appropriate classes to read the data?
Alternatively, you can try using the XSD.exe tool in the .NET framework to have it guess at a schema (the exact parameter escapes me ATM, but it's there).
Since it looks like you'll need to do some middle-tier processing anyway (to break-up the multiple documents in the stream), can you load up each document into an XPathDocument and parse out the needed info? You could then use your favorite ADO.Net technique
to load the data...