Production AI data platform for healthcare — 55,500+ encounter records through a 5-layer trust stack: ingestion (B1), truth contracts (B2), semantic knowledge products (B3), self-healing reliability (B4), and AI spend governance with Vertex Context Caching (B5). Zero data-quality violations reached AI-facing endpoints across 1,000 fault-injection runs. 59.3% AI inference cost reduction via novelty-driven attention routing.
B1→B5: every record earns its way to Baymax. Bad data is quarantined before AI sees it. Good data gets the right compute budget. Fully auditable.
Tech Stack
Features
B1: Trusted Ingestion
55.5K records, 100% source-to-warehouse reconciliation. Entity resolution to 40,235 canonical patients. Domain-specific plausibility contracts (CLINICAL-002) quarantine impossible records before they reach AI.
B2: Truth Contracts
7 named truth contracts (6 BLOCKING), 48/48 GE checks, 52/52 dbt tests. Zero data-quality violations reached AI-facing endpoints across 1,000 fault-injection runs.
B3: Semantic Knowledge Products
4 versioned semantic profiles (PatientProfile, RiskProfile, MedicationProfile, EncounterSummary). BM25 RAG retrieval: Hit@5=0.95, MRR=0.90, NDCG@10=0.89. Grounded answers verified by Vertex Gen AI Eval.
B4: Self-Healing Reliability
99.0% pipeline success, 90% auto-recovery rate, 99.9% SLA compliance. 9-class failure taxonomy. Stale data incidents = 0 across 1,000 seeded fault-injection runs.
B5: AI Spend Governance
novelty_score (text-embedding-004 kNN) drives 89% of PRO-tier routing decisions. Vertex Context Caching: 61.7% per-call cost reduction (measured via count_tokens API). 59.3% cost savings vs naive all-Pro routing across 401 encounters.