It sounds so boring but it turns out that every state has their own process (that they change without notifying you) and they use different pieces of data to identify workers and companies. Also everything was still paper-based so we got a scanned image and OCR text.
This was 10 years ago and we were doing fairly large scale NLP in PHP, matching millions of documents to claimants and companies and doing some amount of fraud detection.
It's still among the most interesting projects I've worked on.