2024
Criminal Records OCR
Desktop application built with the Austrian Interior Ministry's innovation unit that extracts data from large volumes of criminal records and aggregates it into a structured spreadsheet.
The problem
Analysis work was starting from stacks of paper records. Getting anything out of them meant someone reading each one and typing the fields into a spreadsheet. That is slow, and any typo becomes invisible the moment it is in the sheet.
The approach
It had to be a desktop application rather than a web service, because the records were not allowed to leave the machine they were already on.
An Electron shell with a React interface drives an internal Python backend that runs the OCR. Off-the-shelf recognition did badly on the layout and print quality of these particular documents, so I tuned the pipeline against them. The results come out as an Excel file in the format the analysts were already using.
The outcome
A manual transcription job over a large set of records became a batch process, with output in the format the downstream analysis already expected.