The idea
Placeholder: a conversational assistant grounded in a curated body of expert knowledge — lectures, workshops, and written material — rather than generic web-scale data.
The problem
Placeholder: expert knowledge is trapped in recordings and documents. Finding the right answer means re-watching hours of material or asking someone directly.
Approach
Placeholder: retrieval-augmented generation over a structured knowledge base, with careful attention to chunking, context design, and answer attribution.
Current state
Placeholder: actively in development. Core retrieval pipeline works; evaluating answer quality and grounding.
Open questions
- Placeholder: how to keep the knowledge base current without manual curation?
- Placeholder: what is the right balance between faithfulness and fluency?