I started in healthcare. The domain changed; the work and the ideas didn't.
Six years building claims, member and provider analytics at a Blue Cross Blue Shield plan. Then enterprise warehousing at one of the largest US staffing companies, and now leading a data platform team at Magnit Global. The same problems kept coming back: one trusted definition, one identity per member or supplier, governed data that leaders and AI can both use.
-
2010 – 2016
CareFirst BlueCross BlueShield
Programmer Analyst II · healthcare payer
Claims, member, eligibility, provider and clinical data from NASCO, Facets, EHR-related systems and provider applications. ICD-9/ICD-10 coding, PCMH analytics, and PHI handled under HIPAA.
-
2016 – 2019
Allegis Group
Senior Data Engineer / DW Architect
Enterprise warehouse with a bus matrix and conformed dimensions across business units; Informatica ETL across 10+ source systems; MDM golden-record and data-quality patterns.
-
2019 – today
Magnit Global
Engineering Manager, Data Analytics & AI
A 9-person team and a platform serving 110+ enterprise clients. Snowflake platform live in production since May 2025; Power BI semantic layer with row-level security; AI on governed KPIs.
-
Next
Back to healthcare
With a decade more of platform leadership
Idea 01 and Idea 02 on this page: the patterns I've built since, applied to cost of care, population health and payer analytics.
Delivered What I've actually led
The same thread through every role Delivered
| What the ideas need | CareFirst · healthcare | Allegis | Magnit Global · today |
|---|---|---|---|
| AI on governed metricsIdea 01 | SSRS and SSAS analytics used directly by healthcare leadership for operational and strategic decisions. | KPIs and analytical models defined with business stakeholders and adopted across divisions. | Snowflake Cortex pilot: analysts ask plain-English questions against governed KPI views, with accuracy evaluated before wider rollout. |
| One definition, every toolIdea 02 | Type-2 SCD dimensional models for claims, member and provider data, used for historical trend analysis. | Data Warehouse Bus Matrix with conformed dimensions and facts across business units. | Lean Data Model: reusable business definitions and thin semantic models across 5+ VMS platforms, cutting onboarding effort ~40%. Centralized Power BI semantic layer replaced per-client reports. |
| Member & provider identityMember 360 · Provider 360 | Integrated NASCO, Facets, EHR-related and provider systems into central data environments. | Informatica MDM matching and golden-record patterns. | AI-assisted MDM: supplier matching, hierarchy detection and metadata enrichment across systems. |
| Governed, trusted platformPHI · lineage · release control | PHI handled under HIPAA with access control, reconciliation and auditing from development to production. | Data-quality program with Informatica IDQ; CI/CD for warehouse releases. | Snowflake Bronze/Silver/Gold with Kafka CDC and Fivetran ingestion, dbt, Airflow, row-level security and release governance. Oracle → PostgreSQL migration of 100+ reports with zero client disruption. |
Four capabilities, one decision system
Every use case on this page runs through the same loop. Pick a stage to see the question it answers.
Descriptive and diagnostic analytics explain cost, utilization, population, provider and clinical patterns. Statistical methods separate real movement from noise.
Models help prioritize; they do not replace judgment. Risk stratification, forecasting and propensity surface emerging populations and likely future utilization.
AI sits on top of governed analytics. It explains evidence, supports plain-language investigation and suggests the next analysis, while people stay responsible for decisions.
The foundation under all four: each metric is defined once, owned and certified, and served the same way to any BI tool, notebook, app or AI agent.
The loop closes by measuring intervention outcomes against a baseline across utilization, clinical outcomes and financial performance.
How I'd answer "why did cost of care go up?"
A working demo of the approach, not a product pitch. Pick a question: each answer shows the governed metrics it used, the data checks it ran, and where a person has to decide. It builds on what I've already started at Magnit with a Snowflake Cortex pilot over governed KPI views.
Baseline → intervention → outcome → attribution → learn. Synthetic figures; the method is the point.
Build the capability, not another dashboard
Start with the decision leaders need to make. Build reusable analytics and data products around it. Use predictive models where they improve prioritization, and AI where it makes governed evidence easier to explore. Then connect the insight to an accountable workflow and measure whether it created value.
Start with the decision to improve, then map existing or new technology to it.
Use the strategic platforms the organization already pays for when they meet the need.
Trusted member, provider, cost, quality and authorization products serve many consumers.
Every analytics or AI use case has an owner, a baseline, a target and a way to measure it.
Listen and inventory
- Meet finance, actuarial, clinical, network and pharmacy leads
- Inventory platforms, data assets, metric definitions and skills
- Find where PMPM numbers disagree today
Agree the model
- One cost-driver taxonomy, signed off by finance and actuarial
- Governed PMPM, utilization and unit-cost definitions
- Pick one high-value driver path, such as ER or specialty Rx
Prove one thin slice
- Signal → decomposition → cohort → owner, on existing BI
- Conversational investigator on certified metrics only
- Baseline set so realized value can be measured
The detail behind the pitch
Everything above is the short version. Below is the thinking in depth, for when an interview gets to "how exactly would this work?"
An intelligence layer on the reporting you already own
Health plans already have warehouses, semantic models and dashboards. What they lack is a fast, trusted path from "this number moved" to "here is why, here is what we can influence, and here is who owns it." The layer below adds that path without replacing a single platform.
Sources & platform
- Claims, pharmacy, eligibility
- Provider, contracts, prior auth
- EHR / FHIR, labs, SDOH
- Snowflake, Databricks, Fabric, BigQuery
Semantic model & BI
- Certified PMPM, utilization, unit cost
- Power BI, Tableau, Looker
- dbt / semantic-layer metrics
- Row-level security
Conversational intelligence
- Understands the question
- Uses only governed metrics
- Decomposes the trend, cites evidence
- Flags data-quality gaps
- Suggests the next analysis and owner
Decision makers
- CFO, actuarial, finance
- Analysts (hours back)
- Care & utilization management
- Network and pharmacy leads
- Member navigation teams
Members
- Earlier help. Rising-risk members are found while a follow-up visit can still prevent an ER trip.
- Lower out-of-pocket. Navigation toward appropriate, lower-cost sites of care.
- Faster answers. Prior authorization treated as a measured journey, with turnaround tracked.
The health plan
- Trend explained in minutes, not a two-week analyst cycle.
- Opportunity separated from savings, so finance trusts the number.
- AI value that is measured, with a baseline and owner per use case.
- Interoperability readiness for the CMS 2027 API requirements.
Providers
- Shared facts for practice-pattern and value-based contract conversations.
- Less authorization friction through clear requirements and reasons.
- Fewer surprises at reconciliation, because cost and quality are read together.
One trusted number for PMPM, utilization, quality and outcomes, in every tool.
Technically: a platform-independent semantic layer
Today a health plan's most important definitions, such as PMPM, utilization per 1,000 and HEDIS rates, are locked inside individual BI files. Each tool computes its own version, every migration means rewriting them, and AI can't read them. A Trusted Metrics Foundation keeps the business logic in one governed place that any warehouse can run and any analytics tool can query.
Six teams ask the same question: "What was commercial allowed PMPM last year?"
- Snowflake
- Databricks
- BigQuery
- Microsoft Fabric
- Redshift · Postgres
- Metrics and dimensions as code, versioned in git
- Compiled to each platform's own SQL
- PHI and row-level policies defined once
- Certification, owners and lineage
- Caching for speed and cost
- BI: Power BI, Tableau, Looker
- Excel and Sheets
- Python and notebooks
- Apps via REST / GraphQL
- AI agents via MCP
metric:
name: allowed_pmpm
label: Allowed PMPM
owner: finance-actuarial
status: certified
version: 3.2
type: ratio
numerator: total_allowed_amount # incurred basis, IBNR-completed
denominator: member_months
filter: line_of_business = 'commercial'
time: incurred_month
dimensions: [market, product, service_category,
provider_group, cohort]
policies:
phi: aggregate_only # no member rows in BI
min_cell_size: 11 # small-cell suppressionStart with the metrics payers argue about most
Each one gets an owner, a certification status, a version history and a test suite. When actuarial changes the IBNR method, every dashboard, extract and AI answer changes with it the same day.
Switching BI tools or warehouses means repointing connections, not rewriting hundreds of measures.
The board deck, the actuarial filing, the dashboard and the AI answer all agree.
Agents read certified definitions instead of guessing from table names. Idea 01's investigator runs on this foundation.
Access rules and small-cell suppression travel with the metric, whatever tool asks for it.
- InventoryFind every place PMPM and its siblings are defined, and where the numbers disagree.
- CodifyDefine the top 20 healthcare metrics as code with finance, actuarial and quality sign-off.
- ConnectPoint two BI tools and the AI investigator at the layer; reconcile against today's reports.
- RetireRemove duplicate logic from BI files as each team moves over.
Built on open semantic-layer standards, so the foundation itself is not a new lock-in: the Open Semantic Interchange (OSI) spec, published under Apache 2.0 with Snowflake, Salesforce, dbt Labs and Databricks among its backers, and MetricFlow, open-sourced under Apache 2.0 in October 2025.
Start with a business question
Each use case applies the same model to a different decision. Cost of care is the flagship proof of concept; population health is the second deep dive.
What payers are up against in 2026–2027
Affordability pressure is coming from several directions at once. Isolated reports cannot keep up with that; a shared model that decomposes cost and connects it to action can.
Projected 2027 group medical cost trend
8.5% for the individual market; the highest in nearly two decades. Survey of 27 health plans covering 103M+ employer-sponsored members.
of plans rank provider AI coding tools a top-3 inflator
Providers are using AI to optimize revenue. Payers need intelligence that can see the effect in their own claims.
growth in behavioral-health claims utilization
From 2018 to 2024, listed among the major inflators alongside pharmacy and reimbursement pressure.
expect GLP-1 and specialty-drug margin impact / feel well prepared
Health-plan finance leaders see the pharmacy problem coming but lack the capability to manage it.
of AI scalers consistently measure AI's financial impact
44% of organizations are scaling AI, yet few can show what it returned in revenue or cost savings.
CMS prior authorization and access APIs due
Impacted payers must support Patient Access, Provider Access, Payer-to-Payer and Prior Authorization APIs under CMS-0057-F.
A 9% trend does not mean 9% is addressable. The capability that matters is separating utilization, unit cost, population mix and service mix, locating where each is concentrated, and proving which interventions changed the outcome.
Cost driver explorer
Click any branch. The tree breaks an 8.2% PMPM increase into the categories that caused it, measured in percentage points of total growth, then shows whether each one is about more care, pricier care, or a different population.
Utilization
How often care happens: visits, admissions, procedures, scripts per 1,000 members. Is it clinically expected, access-related, or potentially avoidable?
Unit cost
Allowed cost per visit, admission, procedure or script. Reimbursement, facility mix, network status and contract changes all show up here.
Population / risk mix
Age, morbidity, geography and product mix. Separating this stops leaders from treating every increase as an operational failure.
Service mix
What care is delivered and where: specialty drugs, complex procedures, or a shift from office (POS 11) to hospital outpatient (POS 22).
Cost is a journey, not a row in a claims table
The same rising ER line in the tree is made of real member journeys. Seeing them that way shows where an earlier, cheaper, better intervention was possible.
- SignalPMPM, utilization or quality moves outside its expected range.
- InvestigateLocate it across medical, pharmacy, market, product and population.
- DecomposeSplit utilization, unit cost, risk mix, service mix and network effects.
- Root causeTrace to service, provider, facility, condition, drug or cohort.
- AddressabilityExpected, appropriate, influenceable, or potentially avoidable?
- InterveneRoute to clinical, network, pharmacy, member or finance owners.
- MeasureExpected vs actual cost, quality, utilization and member outcomes.
- LearnScale, redesign or stop; refine rules and models.
Check claims completeness and IBNR, membership shifts, seasonality and metric definitions before escalating.
Separate clinically appropriate or structural cost from what the plan can realistically influence.
Report realized impact against a baseline, never identified opportunity as savings. Read cost with quality.
From a whole membership to the members a program can help
Turn a broad population into a clinically and operationally meaningful priority cohort, then measure whether intervention changed utilization, outcomes and cost. Click a stage.
Why is this cohort emerging?
Analytics has to continue after the prediction
- ObserveLongitudinal member signals
- StratifyRisk and cohort analytics
- ValidateClinical and operational context
- InterveneHuman-led program workflow
- MeasureClinical, utilization and financial outcomes
The investigator can summarize governed cohort evidence, compare segments and suggest the next analytical question, for example "Why are these members moving into rising risk?" Program eligibility and clinical decisions stay with people.
All member counts are synthetic and use the same 117.4K-member plan as the rest of this page. This shows an analytics operating model, not clinical decision-making.
What-if intervention simulator
Test an intervention hypothesis against its addressable opportunity before it becomes a business case. Move the sliders.
Scenario
Modeled result
Identified opportunity is not realized savings. Synthetic figures.
Capability first, vendor second
The business model, data products and metric definitions stay stable. The technology underneath can change with strategy, existing investment, skills and cost. Reuse strategic platforms; add technology only where there is a real gap.
- Data sourcesClaims · pharmacy · eligibility · provider · EHR · labs · prior auth · SDOH · contracts
- Integration & interoperabilityBatch · APIs · FHIR · HL7 · X12 EDI · CDC · streaming · data quality
- Data platformLakehouse · warehouse · streaming · metadata · security
- Healthcare data productsMember · provider · claims · pharmacy · cost · quality · authorization · network
- Semantic & intelligenceGoverned metrics · ontology · reusable models · APIs · Trusted Metrics Foundation
- AI & advanced analyticsForecasting · anomaly detection · GenAI investigator · agents
- Decisions & workflowsCare · provider · network · pharmacy · prior auth · outreach · finance
- Outcomes & learningCost · quality · experience · ROI · feedback to models
Member 360
Identity, eligibility, risk, conditions, utilization, cost, care gaps, engagement.
Provider 360
Identity, specialty, network, contract, cost, quality, referrals, outcomes.
Claims & cost
Claim lines, allowed and paid, PMPM, utilization, episodes, attribution.
Pharmacy
Drug spend, specialty, therapeutic class, adherence, prescribers, site of care.
Quality
Measures, care gaps, preventive care, readmissions, outcomes.
Authorization
Requests, decisions, turnaround, reasons, appeals, downstream utilization.
Network & contracts
Status, leakage, reimbursement, fee schedules, contract terms.
Risk & cohorts
Stratification, disease cohorts, emerging risk, intervention populations.
Technology options by ecosystem
| Capability | Microsoft | AWS | Independent / hybrid | |
|---|---|---|---|---|
| Storage | ADLS / OneLake | S3 | Cloud Storage | Object storage |
| Data platform | Fabric / Databricks | Redshift / Databricks | BigQuery | Snowflake / Databricks |
| Transformation | Fabric / dbt | Glue / dbt | Dataform / dbt | dbt / Spark |
| Orchestration | Fabric / ADF | MWAA / Step Functions | Composer | Airflow / Dagster |
| Streaming | Event Hubs | Kinesis / MSK | Pub/Sub | Kafka / Confluent |
| BI & semantic | Power BI | QuickSight | Looker | Tableau / dbt semantic layer |
| AI / ML | Azure AI | SageMaker / Bedrock | Vertex AI | Databricks / model ecosystem |
| Governance | Purview | Lake Formation | Dataplex | Collibra / Alation |
| MDM / identity | Enterprise choice based on the existing ecosystem, e.g. Informatica, Reltio or current MDM | |||
Claims history, financial reconciliation, trend reporting, quality calculations. Daily latency is fine.
Member and provider apps, authorization workflows, FHIR access APIs.
Admissions, discharges, authorization status changes: events where faster awareness changes a decision.