The first thing I built that answered questions instead of storing them.
Two brands, one body of knowledge
Marketing operations ran across two brands with overlapping but not identical rules. Fare conditions, tone, partner terms, which claim may be made where. The information existed. It sat in documents that people had stopped opening because finding the right paragraph took longer than asking a colleague.
Asking a colleague is a retrieval system with one node and a queue.
What I actually had to solve
The technology was the easy part. Three things were not.
Which brand is asking. The same question has two correct answers depending on which brand the answer is for. Retrieval had to know the context before it searched, not filter afterwards.
What counts as a source. A knowledge base is only as good as the decision about what goes in. Half the work was deciding what to exclude. Superseded documents that still read as authoritative are worse than no document at all.
How wrong answers surface. A system that always answers confidently trains people to stop checking. That was the design problem, and it is the one I carried into everything I built afterwards.
Why this one matters
It is the smallest project on this list and the one that changed the most. It proved to me that the barrier to building was no longer engineering capacity but knowing which problem was worth solving. Everything in terminal grew out of what this system got wrong first.