Skip to content

Kevin Holland. Make messy healthcare data usable for products and AI.

I’m open to a full-time Head of Data and AI role, and I take fractional or embedded work for healthcare data businesses. At Holmusk, through October 2026, I owned roadmap, pricing and hiring for NeuroBlu’s real-world data platform, and grew my team from one product manager to six reports across product management, product marketing, analysis and data science.

See the case study

Is this you?

  • An AI data product has to ship, and no one owns it.

    Why now. The pilot has dates. The product leader seat is empty, so decisions wait on the acting CEO.

    Why the usual options fall short. A full-time search takes months. The pilot can’t wait, and the CEO can’t also run the product.

    Hire me as your Head of Data and AI, or bring me in until you do.

  • Your team can’t serve every client.

    Why now. Interpretation is so manual the team can’t scale to every client. The current tool is stuck, handed from one person to the next.

    Why the usual options fall short. The vendor tool checks the box, but it isn’t what you want. Privacy rules block an AI assistant on patient data. A new hire still lacks the time to build.

    A fractional or embedded project, built with your team. To start small, there’s a five-day, fixed-fee Agent Production Gate.

How I work

  1. ClarifyName the decision and the owner.
  2. GroundTest the claim against real evidence.
  3. BuildShip the smallest system that can change the call.
  4. OperateMeasure the outcome and close the loop.

The record

  1. HCA

    Data engineer, then product manager

    Shipped to 175 hospitals and 1,500 leaders. In-house delivery and a renegotiated vendor contract saved about $1.2M a year in reporting costs.

  2. Revive

    Sr. Director, Data Products & Analytics

    Grew a three-product health-system data line to $1M ARR.

  3. Holmusk

    Director of Product, NeuroBlu, 2024 to 2026

    Led product work across NeuroBlu’s $15M ARR platform business, including $3M in standalone subscriptions. Coached the product lead for NLP-derived data products; the team’s label sets grew 50% over two years.

Brand analysis rebuilt for every request became a configured pipeline.

At Holmusk, January through June 2026, brand analysis that took four to eight weeks of custom work per brand became one configured pipeline.

That case is not end-to-end, and no staffing saving or revenue was measured.

About 45 seconds per configured drug from the initial pipeline. That is initial-pipeline output, not an end-to-end rebuild time.

Read the case study

  • CensusChat architecture: a plain-English question enters an MCP validation layer (table and column allowlists, row caps, dangerous-pattern blocking) before reaching DuckDB. Output is a sortable table and chart, under two seconds. Anything that fails the gate never reaches DuckDB. CensusChat A text-to-SQL demo is easy. A text-to-SQL product a healthcare org will point at data is a security problem wearing a chat interface. Input A question in plain English “How many adults over 65 in Cook County?” The key decision: MCP validation layer Generated SQL is never trusted Table and column allowlists Row caps Dangerous-pattern blocking Anything that fails the gate never reaches DuckDB Output A sortable table and a chart, under two seconds 3,144 counties · 239,741 block groups · 84 variables CensusChat architecture: a plain-English question enters an MCP validation layer (table and column allowlists, row caps, dangerous-pattern blocking) before reaching DuckDB. Output is a sortable table and chart, under two seconds. Anything that fails the gate never reaches DuckDB. CensusChat A text-to-SQL demo is easy. A text-to-SQL product a healthcare org will point at data is a security problem wearing a chat interface. Input A question in plain English “How many adults over 65 in Cook County?” The key decision: MCP validation layer Generated SQL is never trusted Table and column allowlists Row caps Dangerous-pattern blocking Anything that fails the gate never reaches DuckDB Output A sortable table and a chart, under two seconds 3,144 counties · 239,741 block groups · 84 variables

    CensusChat

    A natural-language interface over US Census data, with a validation gate between generated queries and the database.

    Open source

  • Data Team OS

    Skills and a file-based workspace for data-product decisions, evidence, quality checks, and human review.

    Open source

  • InstaDoc home page: clinician-governed care online, with three options: a free AI symptom assessment reviewed by licensed professionals, a sick note, and an online prescription.

    InstaDoc.health

    Co-founded with a physician. The live product: symptom assessment and care routing, with clinician review when care is requested. I built the patient-app integration and the synthetic evaluation tools.

    Live product

Writing

All writing

Bring the role or the project

Hiring a Head of Data and AI, or holding an AI pilot with no owner? Send two lines on what it is and when it has to happen.

kevin@kevintholland.com