Documents, extraction, and drift
Real documents rarely arrive in one tidy format. I work on extracting information that another system can use, while keeping enough of a connection to the source to check it.
Reading what is there
Forms, tables, prose, and scans need different handling. A missing field and a confidently invented field are different failures, and both need to be visible.
I want extracted values to be traceable to the document they came from, especially when someone needs to review an uncertain result.
Keeping up with change
Document layouts and model behavior change over time. Trusted examples and schema checks help reveal when an extraction has started behaving differently.
Open questions
- Can every extracted field be traced to a span in the source?
- What would show that today's output no longer matches last month's definition?
- Which failures are missing data, and which are invented data?