A National Hospital Price-Transparency Benchmark, Built on Deterministic Rules and Audit-Ready Controls

Challenge

The company wanted to benchmark hospital prices nationally against Medicare, but the raw material (machine-readable price-transparency files) is large, inconsistent, and easy to misread. Pricing figures shown to employers have to survive scrutiny, so the methodology behind each number had to be documented and reproducible rather than the product of an opaque model. At the same time, the company needed SOC 2 Type 1 to earn enterprise trust, and two of its scoring and statistics engines were due for modernization. The work that had to hold up was the pricing logic itself: every filter, normalization, and benchmark had to be defensible to an enterprise buyer and an auditor alike.

Approach

We built a three-phase national batch pipeline and ran the SOC 2 Type 1 program end to end, keeping AI confined to extraction and language while the pricing decisions ran as deterministic, versioned, documented code.

• Structured the pipeline in three phases: extract per-state from a 165M-row machine-readable-files dataset, aggregate to a compact JSON layer, and generate instant HTML and JSON reports.
• Encoded the pricing decisions as deterministic rules and tables, not model inference: payer normalization, market-suppression thresholds, facility-type exclusions, emergency and professional-fee handling, and invalid-state-code recovery.
• Expressed rates as a percentage of Medicare using CMS MS-DRG weights, so each benchmark traces back to a published, auditable reference rather than a black box.
• Used AI for the reading and the writing (parsing messy inputs, normalizing labels, and drafting methodology and integration documentation), then let the deterministic code own every number that reaches a report.
• Packaged the pipeline for the client’s own cloud and handed it off with integration and methodology docs so the team could run, audit, and extend it.
• Led the SOC 2 Type 1 program end to end: vendor selection, Trust Services Criteria scoping, gap analysis, security policy authoring, and cloud-integration coordination to an audit-ready state.

Impact

• Deployed nationally across 42 viable states, 194 markets, and roughly 2,450 hospitals, producing about 11.6M pricing records after data-quality filtering.
• Benchmarks are expressed as a percentage of Medicare via CMS MS-DRG weights, with the methodology documented so each figure is reproducible and defensible to enterprise buyers.
• Data-quality rules (payer normalization, market-suppression thresholds, facility-type exclusions, emergency and professional-fee handling, invalid-state-code recovery) run deterministically, so results are consistent run to run and explainable line by line.
• The SOC 2 Type 1 program reached audit-ready, including 15 security policies authored across versions, scoped Trust Services Criteria, and completed gap analysis. (Type 1 attests to control design at a point in time; it is the readiness milestone, not an ongoing operating-effectiveness opinion.)
• The pipeline was packaged and handed off to run in the client’s own cloud, with integration and methodology documentation, reducing dependence on the original build environment.
• Separately modernized a benefits-quality scoring engine and an automated workforce-statistics pipeline sourced from public APIs.

Capabilities demonstrated

• Large-scale healthcare data engineering against multi-state machine-readable price-transparency files (165M-row source).
• Deterministic, auditable benchmarking logic with rates anchored to CMS MS-DRG weights.
• Rigorous data-quality design: normalization, suppression thresholds, exclusions, and error recovery as documented rules.
• AI used narrowly for extraction and documentation, with all decisions running as versioned code.
• SOC 2 Type 1 program leadership across vendor selection, TSC scoping, gap analysis, and policy authoring.
• Clean handoff to a client cloud with integration and methodology documentation.

Anonymized by design: client names stay off the narrative per our reference policy. Figures that are modeled, small-sample, or targets are identified as such above.