Building healthcare analytics that prove their own numbers. Right now that means:
- Claims anomaly detection with reason codes: hybrid rules + isolation forest, recall 1.00 on 4,643 seeded anomalies, 21 ms scoring
- Data quality as a CI gate, not a dashboard: YAML contracts + learned baselines at 390k rows/s; caught a 24% mean shift that PSI scored a harmless 0.047
- Forecasts that must beat seasonal naive to ship: when naive won on 13 of 40 SKUs, naive kept the crown
- Causal inference calibrated on null panels first: naive math showed 16.5% false positives; the shipped rule holds 11.5% at a nominal 10%
My career has one hinge in the middle. I started at McKesson in India, close to pharmacy and claims operations, where I learned that a number nobody trusts is a number nobody uses. That question (how do you make analytics trustworthy?) pulled me to the UK for an MSc in Data Science at Teesside University, and then into Optum UK, where the datasets got bigger and the stakes got clinical. The through-line never changed: healthcare data, and the engineering discipline to make it hold up under audit. Every repo below is that discipline in miniature: seeded ground truth, reconciliation invariants, calibration before claims.
|
languages + analysis
|
data platforms
|
|
BI + storytelling
|
cloud + delivery
|
| Repo | What it proves |
|---|---|
| claims-anomaly-radar | Hybrid rules + isolation forest flags claim anomalies with reason codes; recall 1.00 on seeded ground truth |
| claims-lakehouse-pipeline | PySpark medallion where every layer reconciles to the previous; $308M conserved to the penny, 2M rows under 2 minutes |
| warehouse-cost-attributor | SQL credit attribution that reconciles to the bill within 0.0006%; found 25.4% idle spend |
| pharmacy-demand-forecaster | Rolling-origin backtests where models must beat naive; +1.8pt service level where they did |
| intervention-impact-analyzer | Counterfactual + placebo inference, self-calibrated on null panels before reporting anything |
| sql-regression-radar | CI gate on EXPLAIN plan fingerprints; catches a dropped index with a plan diff, not a hunch |
six more where those came from
| Repo | One-liner |
|---|---|
| healthcare-data-quality-gate | YAML contracts + learned baselines gate pipelines in CI; 390k rows/s |
| scd2-sentinel | Nine SQL checks for SCD2 dimensions; 1.5M-row dim validated in 13.6s |
| kpi-drift-rootcauser | Rate/mix decomposition of KPI moves; contributions reconcile to 1e-9 |
| patient-feedback-miner | NMF topics + budgeted sentiment; 0.982 F1 from 1,000 labels |
| fhir-ingest-gateway | Validating, idempotent FHIR ingestion with DLQ; ~600 resources/s |
| provider-record-linker | Record linkage with measured blocking; 0.978 precision, 0/20 trap merges |
"All models are wrong, but some are useful." - George E. P. Box