Permission-Aware Knowledge Assistant

The answers people needed were spread across a dozen internal systems, and finding one meant already knowing which system to look in. An assistant that could read all of them was the obvious fix and the obvious risk: the moment a single thing can see everything, it can tell anyone anything. Answering questions well was never the hard part. Making certain the assistant could not surface a document the person asking was not allowed to open was.

An assistant that can read everything is only useful if it can be trusted never to tell the wrong person the wrong thing.

Sole engineer, end to end. A Next.js front end, a Python service behind it that finds the relevant documents and talks to the model, and the Azure infrastructure both run on. I designed the permission model, wrote the ingestion that keeps the searchable copy current as the source systems change, and shipped it. The specifics of the internal design are my employer's to share, not mine.