Eric Grynspan

New York, NY
Data Engineer — Healthcare AI & Fintech

I build production-grade data infrastructure and real-time AI systems for healthcare. Current focus: pre-submission RCM prevention — Kafka streaming, LLM tool-use, and agentic denial scoring at the point of submission. Prior work: FHIR R4 claims analytics and LLM governance. Tested, documented, and fully auditable.

Stack
Python SQL Snowflake dbt Dagster Airflow Apache Kafka AWS Terraform FHIR R4 RCM LLM Tool-Use LLM Enrichment Docker
Work
Cleared: Agentic RCM Prevention Pipeline Live
Kafka → LLM tool-use → Snowflake → dbt. ~85% of claims cleared without an LLM call — deterministic NCCI gate, sub-ms. 10% holdout arm for provable lift. Tiered action router: 3-condition auto-correct gate (FCA defense) → immutable audit log → kill-switch. Feedback loop compares intervention vs. holdout denial rates; >20% drift activates kill-switch. $0.003/claim · 50,000× ROI. 251 tests, CI green.
Trust but Verify: Clinical AI Governance Engine Live
FHIR R4 → Snowflake → dbt → Dagster. 25,958 records · 226 patients · 6 enrichment categories. LLM enrichment cross-validated by LLM-as-Judge + deterministic rules engine. Every record routes to Gold (trusted) or Review (explainable reason, always traceable). Dual-validation: neither LLM nor rules engine alone decides.
Denied: Healthcare Claims Intelligence Pipeline Live
FHIR R4 → Snowflake → dbt → Dagster. 257K denied claims · 51.9% denial rate · $1.2M+ recoverable. CARC-based root-cause classification — systematic vs. documentation failures. T2D+CKD RWE cohort. 12 dbt models, 83 tests, CI green.
Own the Model, Own the Returns: Sharpe Premium Pipeline Live
Airflow → S3 → Athena → Power BI · Terraform IaC. +92% Sharpe ratio premium for AI-native builders vs. integrators. Spearman ρ = +0.800, p ≈ 0.005. 184 pytest tests · SEC EDGAR + FRED + live market data.
LinkedIn GitHub dbt Docs