RAG knowledge assistant
An AI assistant that answers 500+ employees in seconds, and knows when to hand off.
Gen AI Architect · team lead
Employees across HR, IT and operations were waiting hours for answers that already existed somewhere in policies, runbooks and knowledge bases, and the teams answering them were spending their time on repeat questions.
How it works.
Grounded in the bank’s own knowledge
Retrieval-augmented generation answers from internal HR, IT and operations content instead of the model’s general knowledge.
Retrieval built for real questions
Employee questions are matched against the knowledge base so the assistant answers from the passages that actually apply.
Answer or hand off
The assistant resolves what it can (about 65% of queries) and routes the rest to people, so it shortens the queue instead of adding a dead end.
Response time dropped from hours to seconds for 500+ employees, about 65% of queries are handled end to end, and the support teams saved the equivalent of 3 full-time roles.
My role
- Led the team that designed and shipped the assistant
- Owned the RAG architecture across HR, IT and operations content
Stack
- RAG
- LLMs
- Vector search
- Python