Enterprise AI architect and causal scientist in regulated healthcare

Akbar Akbari Esfahani

What actually changed, and how do we know?

I've asked that since I classified geophysical signals at the U.S. Geological Survey. Whether the signal comes from the ground, an outbreak or a health plan's workflow, the question hasn't changed.

  1. Measure before you claim.
  2. Build instruments that live inside the work, not dashboards that watch it from outside.
  3. Keep people accountable for the decisions that matter.
More about what I do

Hi, I'm Akbar.

I chair the AI Governance Committee and serve as enterprise AI architect at Central California Alliance for Health, a health plan, where I run an internal AI product studio and lead a team of PhD data scientists. I still write a lot of the code.

Before that I led data science and analytics at Vital Decisions, led data science innovation at L.A. Care Health Plan, and stood up the first AI function at Highmark. I started out as a mathematician at the U.S. Geological Survey and a statistician at UCLA, and I've taught data science at Carnegie Mellon.

Away from work, I'm a husband, a father of two, a sibling of three and a child of tradition.

If you'd like to learn more about what I do:

  1. How I lead: AI strategy, architecture and governance as an infinite game
  2. Causal evaluation, program evaluation and statistical ROI
  3. Building AI governance that can say who approved what, and why
  4. Research: long memory, signals under uncertainty, instruments
  5. What I ship: papers, talks and code
  6. Teaching, and growing from analyst to architect
  7. Talks and speaking
  8. Where I come from: Iran, Munich, three languages

01How I lead

Leadership is an infinite game.

A quarter ends; the organization doesn't. I lead for the long game: set the strategy, draw the architecture, chair the governance and still ship the systems, so the people and the platform can keep playing after any one project, or any one leader, is done.

How I play it

  • A cause bigger than the quarter

    One hour of organizational work should equal one hour of service to the person the organization exists to serve. Measure against that, not against the dashboard.

  • Teams that can tell the truth

    People name what is broken in their work only when naming it won't be held against them. I build that permission first, then the tools.

  • Fix the work before you automate it

    Redesign the workflow, then add models. In that order, my team got 6x staff output and cut median resolution from 56 to 20 days.

  • Build what outlasts you

    Shared platforms, reusable scaffolding and a governance charter the next leader can run. Retire what no longer serves.

  • The courage to say no

    Say what the evidence cannot support, and write the ethical guidelines before anyone asks. I wrote my first set in 2018.

Selected outcomes, delivered with my teams

$26M+
measured annualized value from workflow redesign, causal evaluation, internal software and build-versus-buy decisions
6x
staff output after redesigning a workflow before adding ML and agentic routing; median resolution from 56 to 20 days
Zero
external PHI transmission across multiple production-deployed LLM systems
$675K
in planned hires avoided by fixing the work before automating it

The arc

  1. 2023 to now

    Chair, AI Governance Committee · Enterprise AI Architect

    Central California Alliance for Health

    Manager, Advanced Analytics. Runs an internal AI product studio for a regulated health plan, where departments are the clients. Authored the enterprise AI strategy, governance charter, 24-month roadmap and five-level maturity model. Partners with the COO, Chief Compliance Officer, CMO and CIO on portfolio decisions. Recruited and leads a four-person PhD team and stays hands-on in architecture and production code.

  2. 2022 to 2023

    Head of Data Science and Analytics

    Vital Decisions, acquired by Evolent Health

    Built the analytics function and program-evaluation framework at a B2B health SaaS company serving national health plans, partnered with the CEO on the product roadmap, and carried the data roadmap through the acquisition.

  3. 2021 to 2022

    AI and Data Science Consultant

    AgreeYa Solutions, for the California Department of Health Care Services

    Advised state health agencies on analytics strategy, infrastructure and program measurement for large-scale transformation.

  4. 2019 to 2020

    Director, Data Science Innovation

    L.A. Care Health Plan

    Led data science innovation at the nation's largest publicly operated health plan, with four direct reports. Deployed ML for utilization prediction and 30-day readmission prevention.

  5. 2016 to 2019

    Senior Data Scientist and Head of Innovation

    Highmark Blue Cross Blue Shield

    Stood up the first AI function at a 5M+ member organization: GPU infrastructure, deep learning, production applications and the roadmap. Owned the R platform for 25 data scientists. Wrote the organization's AI and ML ethical guidelines in 2018.

  6. 2010 to 2016

    Mathematician, then Senior Statistician

    U.S. Geological Survey · UCLA Center for Health Policy Research

    Built the USGS's first ML infrastructure for geophysical classification. Led complex survey methodology and population-scale prevalence estimation for the California Health Interview Survey.

02Causal evaluation and ROI

Did it work? Compared with what? Worth what?

Most ROI in healthcare is modeled savings: a favorable trend multiplied by a price. I measure what a program, vendor or AI tool actually caused, put an honest interval on it, and turn only that into value you can defend to a CFO, a board or a regulator.

What I evaluate

  • Programs

    Care management, outreach and utilization programs: did the people who got it do better than comparable people who didn't?

  • Vendors

    Whether a vendor's claimed savings survive a fair comparison, before the renewal and not after.

  • AI and workflow changes

    A before-and-after chart is not an evaluation. The comparison is designed in from the start.

How an evaluation runs

  1. Fix the decision it will inform: continue, modify, expand or stop
  2. Name the effect to estimate and a fair comparison: randomized rollout, matched comparison, difference in differences or interrupted time series
  3. Write the claim rules and refusal conditions before opening the data
  4. Estimate the effect with its uncertainty, and test how fragile it is
  5. Convert only the attributable effect into dollars, net of the full cost

Proper statistical ROI

  • Attributable, not observed

    ROI starts from the causal effect, not from the change over time. Regression to the mean and secular trends are not savings.

  • An interval, not a point

    Report the range the evidence supports and the ROI that range implies, including the chance it is below break-even.

  • Costs in full

    The price of the program plus the cost of running, reviewing and supporting it.

  • Every number labeled

    A count, a comparison or a cause. Only a cause can carry an ROI.

Track record

  • Measured value

    $26M+ in measured annualized value with my teams, through causal evaluation, workflow redesign, internal software and build-versus-buy decisions.

  • Evaluation frameworks

    Built the program-evaluation framework at a health SaaS company serving national health plans.

  • Methods

    Population-scale survey methodology on the California Health Interview Survey at UCLA; first-author peer-reviewed work on forecasting under non-stationarity.

  • A name on the finding

    When a result needs an accountable expert: statistical methodology for an IRB, and de-identification by HIPAA expert determination.

Available for outside consulting: program and vendor evaluations, statistical ROI, IRB methodology and expert determination.

03AI governance

Who approved this, and why?

Everyone wants AI tools. Few organizations can say who approved which one, under what controls, and what happened after it went live. I've built the committee, the controls and the systems that answer that.

What I've built

  • The committee

    Chair of the AI Governance Committee at a regulated health plan, in partnership with the Chief Compliance Officer.

  • The controls

    Four risk tiers scored across nine dimensions, go-live gates, human oversight, audit standards, incident and fallback paths, and retirement criteria.

  • The vendor terms

    Requirements covering BAAs, model training, subprocessors and exit plans.

  • The literacy

    An enterprise AI literacy curriculum, built and delivered with my team.

  • The long view

    AI and ML ethical guidelines written for a 5M+ member insurer in 2018.

  • The public sector

    Advised California's Department of Health Care Services on analytics strategy and program measurement.

04Research

Long memory, and what it takes to observe it.

My research runs from signals in the ground to signals in health systems, and lately to the mathematics of observation itself. The common line: uncertainty as missing information rather than noise, limited measurement that admits many models, and the origin of an event read from the structure of sparse traces.

Threads

  • Long memory and non-stationarity

    Fractal structure in 2,000 years of climate, and conditional climate-change forecasts built from quantile trends and fractionally differenced ARIMA.

  • Signals under uncertainty

    At the USGS: telling unexploded munitions from clutter in electromagnetic data, where meta learners with Bayesian networks replaced a three-week numerical inversion with a model trained and tested in under 30 seconds; and Bayesian McMC interpretation of airborne electromagnetic surveys.

  • The subsurface

    Hybrid machine-learning models for real-time three-dimensional mapping of surficial aquifers.

  • Outbreaks

    Locating the source of the 2011 U.S. listeriosis outbreak from sparse poison-center calls with the Topological Weighted Centroid: by day 26, from 57 calls, the predicted source lay about 94 miles from the farm. And COVID-19 among health care workers: an ARFIMA model of nearly 9,000 tests found a long-memory infection pattern lasting months, with spread among health care workers ahead of the general population.

  • Medical imaging

    Receptive-field features from chest CT with Bayesian networks, to tell cystic fibrosis from healthy lungs.

  • Operational capacity

    What a budgeted observation protocol can and cannot distinguish: bounds on query rank, distinguishable outcomes and history amplitudes in finite models. A working paper, September 2026.

  • In situ analytics

    Observation capacity, not any dataset it produces, is the durable enterprise asset. A conceptual paper, draft of August 2026.

  • Causal program evaluation

    Measurement contracts that fix the question, the comparison and the claim rules before anyone opens the data.

Selected publications

05What I ship

Read what I ship.

The quickest way to see how I think is in the artifacts: code that keeps an answer attached to its source, and instruments that keep a decision attached to its evidence.

Public work

  • Paper

    In situ analytics

    Instrumentation as the unit of enterprise intelligence: the three properties that make a system an instrument, and the cost test for building one versus buying inference. Draft, August 2026.

  • Talk

    Realtime operational instrumentation

    The Ai4 2026 talk: a governed drafting test bed on synthetic records that caught its own citation failure, and was fixed the same day.

  • Open code · lhpc_kb

    California Health Plan Policy Knowledge Base

    A policy knowledge base across all 17 Local Health Plans of California plans, with DHCS All Plan Letters as the spine. Page-anchored retrieval in one DuckDB store, and every chunk names its plan so answers stay with the right one.

In the day job

  • Platform

    Owned-GPU LLM inference, permission-aware retrieval, source grounding, agent orchestration, governed promotion, evaluation, monitoring and fallback paths.

  • Speed with a gate

    Reusable scaffolding that moves a validated use case from prototype to production in under a week.

  • Tools

    Python, R and SQL; PyTorch and scikit-learn; Snowflake, Spark and Posit Connect; on-prem NVIDIA GPUs, Azure and GCP.

06Teaching

From analyst to architect, without losing the science.

I started as a mathematician classifying signals and grew into enterprise architecture. The method came along the whole way, and it is the part worth protecting.

Where I've taught

  • Carnegie Mellon University, Heinz College

    Adjunct faculty, 2018 to 2019. R for data science and data mining for master's students in information systems and public policy.

  • A citizen data scientist program

    Built at Highmark and grown to 100+ people meeting every two weeks. The wider enablement program reached 300+ staff.

  • Enterprise AI literacy

    A curriculum for a health plan's staff, built and delivered with my team.

  • West Virginia University

    "Nuts and Bolts of Being a Data Scientist", a seminar for the biostatistics department.

The path

  1. Mathematician, U.S. Geological Survey
  2. Senior Statistician, UCLA
  3. Senior Data Scientist and Head of Innovation, Highmark
  4. Director, Data Science Innovation, L.A. Care
  5. Head of Data Science and Analytics, Vital Decisions
  6. Chair and Enterprise AI Architect, the Alliance

Three habits worth keeping

  • Keep the method where others can check it.

    A result nobody can reproduce is an opinion with a chart.

  • Ship to the person doing the work.

    Not to a dashboard they have to leave the work to read.

  • Say what the evidence cannot support.

    It is the sentence that earns trust for all the others.

07Speaking

Every talk asks the same question.

For conferences, podcasts and editors looking for someone who builds governed AI in a regulated setting and measures what it changes.

Talks

  • Ai4 2026

    Every dashboard is a dead butterfly

    Realtime operational instrumentation in healthcare: why analytics belongs inside the work, and what an instrument sees that a report never will. Given at Ai4 2026 in Las Vegas.

  • Governance

    Who approved this, and why?

    What an AI governance committee actually decides, the controls that make its decisions checkable, and what happens after go-live.

  • WVU

    Nuts and bolts of being a data scientist

    A seminar for the biostatistics department at West Virginia University's School of Public Health.

  • Measurement

    Count, comparison or cause

    Causal measurement in healthcare: telling a real program effect from a favorable trend, and writing the claim rules before the analysis.

08The person

Three languages, one question.

I'm a husband, a father of two, a sibling of three and a child of tradition. I was born in Iran and grew up in Munich, I think in English, Farsi and German, and the tradition I come from, Persian literature and the Sufi teachings, shaped me as much as statistics did.

English
What actually changed, and how do we know?
Deutsch
Was hat sich wirklich verändert, und woher wissen wir das?
فارسی
چه چیزی واقعاً تغییر کرد، و از کجا می‌دانیم؟

Both traditions ask for the same discipline: attend closely before you judge, and be honest about the limits of what you know.

In systems, that means treating an organization as relationships before tables, and building instruments that sit with the people doing the work. In people, it means assuming the person closest to the work saw the change first.

Start with the question.

Tell me what changed, or what you need to know changed. I read everything that arrives here.