01The challenge
Agency consultants use the application as a retrieval-based AI knowledge base and to monitor legislative and regulatory changes while working on client projects. Their research spans source material that changes over time.
The product needed to connect three questions: what changed, what does the evidence say and where does a person need to look more closely?
02My role
I gathered stakeholder needs, wrote user stories, designed the architecture and frontend, and built the application with AI assistance. I also separated implementation tests from answer-quality evaluation.
04The result
The agency uses the knowledge base and monitoring application in client-project work. No formal answer-quality evaluation has been completed, and no accuracy or time-saving result is claimed.
Development continues on scanned-PDF support, retrieval ranking and source coverage. The recording shows the monitoring module.
What I’d measure next
- Measure
- Answer correctness, source support and appropriate refusal.
- How
- Build a representative, expert-reviewed question set, including unanswerable cases. Score correctness and whether cited sources support each claim. Report unsupported claims, missed refusals, source coverage and time for a reviewer to verify an answer.
- Check alongside it
- Evaluate access-rights failures separately. Deployment and citation matching are not proof of reliable answers.