AURORA turns a rough writing request into a draft accompanied by retrieved sources, quality checks and a trace of how it was produced. We built it as a proof of concept for ABN AMRO with the AI Student Organization (AISO), for ABN AMRO Insights content in Technology, Media and Telecom, with English, Dutch and bilingual output.
I wrote about the project while it ran: why retrieval is the main design decision and the mid-cycle pipeline.
Why we built it
A short prompt rarely contains everything an editor needs: the audience, the relevant source material, the bank’s writing guidance, and the criteria for a usable draft. Supplying those details again and again leaves much of the work with the author and the reviewer.
Our central design decision was to make that preparation part of the system. AURORA interprets the request, selects editorial and domain profiles, retrieves articles and writing guidance, and refines the assignment before any drafting happens. Evaluation then gives the reviewer specific findings next to the text.
What it does
| Step | What happens |
|---|---|
| Interpret the request | Classifies the task, sector, keywords and language, and selects drafter, reviewer or curator profiles. |
| Retrieve sources | Searches English and Dutch articles and writing guides, and returns snippets with their source. |
| Refine the prompt | Adds context and constraints, and asks clarification questions when the request is too open. |
| Generate the draft | Writes the draft with citation references, in English, Dutch or both. |
| Evaluate | Combines deterministic checks, six LLM quality judges and the human reviews still required. |
| Trace | Keeps the stage results, selected sources and findings for every run, so an editor can inspect them. |
The same services are exposed three ways: a REST API built with FastAPI, eight MCP tools, and a Streamlit interface with a guided mode and a pipeline inspector.
The integration design below shows how an approved corporate agent could call these services. Production access controls and durable audit storage remain requirements for a pilot.

How evaluation works
The evaluation catalogue holds 135 KPI entries derived from ABN AMRO’s content KPI workbook. Only the part that applies to a draft runs during generation.
- Deterministic checks catch invalid citation markers, missing image alt text and citations to sources that may not be used.
- Six LLM judges assess factuality, adherence to bank standards, relevance, privacy and security, source groundedness and completeness.
- Human review requirements, such as editorial and legal sign-off, are reported as pending. A passing evaluation is not publication approval.
Status
AURORA is a working proof of concept with 97 passing tests. It shows an inspectable drafting and evaluation workflow. It does not yet show measured time savings or quality gains: that needs a controlled pilot comparing editorial tasks with and without it, which the final report proposes.