GovTech Document Intelligence

Overview

An AI document management platform for public-sector records: staff ask questions in plain language and get answers grounded in the agency's own documents, with the source of every claim attached. Personal data is detected and redacted automatically before anything leaves the system, and a human signs off on each redaction.

The problem

Public-sector document work carries obligations private-sector tooling ignores. An answer without a traceable source is unusable in a process that may be reviewed years later, and personal data has to be removed before a record travels — including to the model itself. Generic assistants were fluent and confidently wrong, which is worse than no assistant at all.

What we built

Retrieval with citations

Answers are assembled from retrieved passages and rendered with their sources inline, so the citation is part of the answer rather than an afterthought.

Automated PII redaction

Detection runs over every document before it enters retrieval, with a mandatory human review step before a redaction is treated as final.

AI-assisted correspondence

Draft emails composed against retrieved records, so the officer edits something grounded rather than starting from a blank page.

Access and audit trail

Every retrieval and every redaction decision is logged against a user, because the question 'who saw this' has to be answerable.

How we approached it

• Traceability beats cleverness — an uncited answer is not an answer here
• Redaction is automated but never unattended
• The system is built to survive audit, not to demo well
• Retrieval is scoped by permission, so the model cannot surface what a user cannot see

Outcome

A production system agency staff use for daily records work, where every answer carries its sources and no personal data leaves the boundary without a reviewed redaction.