That is going a lot of directions
If you have known worker attributes like doctor, engineer, accountant you could have mirror document classification and us ML to classify.
If you are going to measure performance then you would need to need get a sample from each user. Take like the 1/4 and have a target vector for the user. Map a document to a users strength.
Static parameters would be different and I am not sure you should use them if you have performance measure. I guess to queue before you have performance for a worker. But even at 4 documents you have a top 1/4.