A governed medallion lakehouse that ingests messy, multi-source health data and turns it into three trusted outputs — analytics marts, an ML feature store, and a vector index for retrieval.
Many source systems, competing vocabularies, silent unit drift, PHI everywhere, and free text where coded values belong. Left unmodeled, it produces silently-biased analytics and unreliable models.
Bronze → silver → gold, with a healthcare layer overlaid — the part a generic ETL project doesn't have.
Dimensional fct_/dim_ models plus a semantic layer — one metric definition shared by BI, cohorts, and ad-hoc queries. The trusted serving layer.
20 time-windowed features across four source types. Online store for low-latency inference; offline for leakage-safe training joins. Both parity-checked.
390 clinical notes indexed with HNSW cosine, TF-IDF fallback when Docker is down. Retrieval-only — it proves the data is AI-ready.
PHI is tagged and access-gated at bronze; silver is the de-identified boundary — HIPAA Safe Harbor plus per-patient date-shifting to preserve temporal order.
ICD-10, SNOMED CT, LOINC and RxNorm mapped into a recognizable OMOP Common Data Model on the way in.
Validity, completeness, unit consistency, uniqueness and timeliness enforced as contracts at the silver gate — and the gate runs in CI, exiting non-zero on any violation.