Reading members' documents without keeping them
Letting people ask about their own contracts, notices and photos of documents: text is extracted, personal identifiers are masked before any model sees it, and nothing is stored.
Goals
• Extract text from PDF, Word, text files and photos, with Vietnamese OCR, in a temporary folder only.
• Mask ID numbers, tax codes, bank accounts, phone numbers and emails before the text reaches a model, and again on the server.
• Search the library once per clause of the document, so each obligation it mentions is checked against the law.
• Drop the document from saved chat history; show people a preview so they can check what remains (names and addresses aren't masked).
Skills we are looking for
- OCR
- Privacy engineering
- Vietnamese NLP