Data Engineer building audit-ready pipelines for healthcare and finance: deterministic controls for payment integrity, with errors intercepted upstream and every figure traceable to a test. Denied separates systematic denials from documentation failures across 257K claims (modeled $1.2M+ recoverable). Cleared stops denials pre-submission: a rules-based NCCI gate resolves ~85% of claims before any LLM call, at $0.0119 each, against a 10% holdout. Trust but Verify dual-validates AI-enriched records so every routing decision is explainable. Snowflake · dbt · Dagster · Kafka · Airflow · AWS.
Denied classifies denials retrospectively → Cleared prevents them in real time → Trust but Verify governs the AI that touches the records
Classifies 257K denied claims by root cause — systematic denials vs. documentation failures — where the remediation path differs fundamentally for each.
| Stack | Synthea FHIR R4 · Python · Snowflake (RAW → staging → mart) · dbt · Dagster |
| Scale | 495K total claims · 51.9% denial rate · $1.2M+ recoverable · 12 dbt models · 73 tests |
| RWE | T2D/CKD cohort · 104 patients · 54.8% metformin utilization |
Intercepts claims before submission instead of analyzing denials after the fact — a deterministic NCCI gate plus LLM tool-use prevent the two largest denial root causes at the source.
| Stack | Apache Kafka (KRaft) · LLM tool-use · Snowflake · dbt · Python |
| Design | NCCI gate (~85% cleared without LLM) · 3-condition auto-correct gate (confidence ≥ 0.92 + charge ≤ $500) · drift monitor (>20% → kill-switch) |
| Proof | 10% holdout control arm · provable intervention vs. control lift · $0.0119/claim live-validated · ~12,600× ROI · 284 tests · CI green |
Dual-validation AI governance — LLM enrichment cross-validated by a deterministic rules engine. Every record routes to Gold (trusted) or Review (explainable reason). No black-box outputs.
| Stack | FHIR R4 · Python · Snowflake · dbt · Dagster · LLM API · Pydantic |
| Design | LLM-as-Judge blind audit · prompt caching · confidence threshold routing |
| Data | 226 patients · 25,958 clinical records · 6 enrichment categories · 6 rules engine domains |
Proprietary AI builders generate a +92.0% Sharpe ratio premium over third-party integrators (Spearman ρ = +0.800, p ≈ 0.005) across 10 major tech stocks — visualized in an interactive Power BI dashboard.
| Stack | Airflow · S3 · Glue · Athena · Power BI · Terraform · GitHub Actions |
| Storage | Hive-partitioned S3 data lake · Parquet/Snappy · Glue catalog · serverless Athena |
| Quality | 184 pytest unit tests · moto AWS mocking · GitHub Actions CI/CD · Terraform IaC |
