MVP: import one patient report → on-device chunk + embed + index → on-device hybrid retrieval → generation via a controlled endpoint sending only query+snippets → cited answers for a bounded question set (e.g., "explain my pathogenic/likely-pathogenic findings" and pharmacogenomics) → abstain + clinician boundary on everything → metadata-only telemetry.
Defer: multi-report, agentic multi-hop retrieval, fully on-device LLM, fine-tuning.
This cut is shippable, compliant, and demonstrably useful — exactly what to propose when asked "where would you start?"