Regius

Recruiters at Regius start from a résumé and an interview. From those they build a candidate profile they can reuse, and messages written for one specific vacancy. The tool had to let a model write the first version of both, without letting it have the last word.

A FastAPI web app with a background worker pipeline in Python 3.12, in three stages: document extraction, drafting, rendering. Azure Content Understanding, Azure OpenAI and Azure AI Speech do the reading, the writing and the listening. Each case is kept as PostgreSQL snapshots.

The rule that shaped everything else: model output is always an editable draft, and a reviewer’s edit is authoritative. The model never gets to write over a person.

A synthetic evaluation harness: invented candidates and vacancies, so the whole pipeline can be tested end to end without anyone’s real data. Privacy and PII safeguards sit around the pipeline itself.

A working pipeline on Azure, from a résumé and interview material to a drafted candidate profile and vacancy messages, each passing a recruiter before it is used, and tested end to end on synthetic cases.

It runs on Blob Storage, Service Bus and Docker, all of it described in Bicep, so the environment can be rebuilt from code instead of from memory.