The situation
Engineers were spending their time locating information rather than using it. The answers to most of their questions existed somewhere across operational records, technical documentation, and historical datasets, in inconsistent formats.
The constraint
This is exactly the domain where a fluent wrong answer does damage. An assistant that invents a tolerance or misattributes a reading is worse than no assistant, because it is faster and sounds certain. It had to be able to decline.
What I decided
I put the effort into retrieval and grounding rather than the model. Answers are traceable to the document or dataset they came from, and the system is built to say that it does not know when the evidence is not there. Most of the engineering was the data model and the ingestion path; the model itself was the small part, and the replaceable one.
What shipped
Retrieval-augmented assistants over operational and technical data, with contextual reasoning across sources and citations back to the underlying records.
What it produced
Manual retrieval and cross-referencing largely eliminated from the analysis loop, with answers that an engineer could check instead of having to trust.