What I’m Doing Now

Last updated: Monday, May 4, 2026

This is a now page. It changes when I do.

Quick Snapshot

I’m building tools and communities that help patients use AI to understand their own health data, on their own terms. Some of that work is software. Some of it is people. Lately, more of it is software than I expected.

OpenKP: A Patient-Owned Bridge to My Kaiser Record

For the last few weeks I’ve been building OpenKP, a patient-owned MCP server that lets Claude read my Kaiser Permanente medical record locally, with my own credentials, under my control.

The short version of how it started: Kaiser Permanente’s FHIR API doesn’t work for patients. Josh Mandel’s SMART on FHIR skill, the one I used to pull my dad’s Stanford record, hits a CORS wall at Kaiser. So when Ryan Hughes at Fan Pier Labs published Open Record, an MCP bridge to MyChart, I asked Claude whether the same approach could work against kp.org.

Claude said no, not as a fork. Kaiser’s implementation of Epic isn’t vanilla MyChart. But it offered to sketch a Kaiser-native scraper from scratch. I said yes. Three evenings later I had a working prototype that could read my labs, meds, messages, and visit notes locally.

The architecture matters as much as the code. OpenKP runs on my desktop. Nothing leaves my machine. The credentials are mine. The data is mine. Claude answers to me, not to Kaiser. This is the difference between institutional AI and patient-directed AI in one project.

I started this in Claude Cowork. I’ve since moved most of the build into Claude Code CLI, which I had been avoiding because it looked intimidating. It isn’t. It’s the most productive coding environment I’ve ever worked in, and I’m not a developer. That shift, from a chat-shaped tool to a CLI agent, is itself part of the story.

Reading my dad’s record at Stanford

In parallel, I’ve kept analyzing my 95-year-old dad’s Stanford record using Josh Mandel’s Health Skillz with Claude. Fifteen years of care, hundreds of encounters, structured FHIR JSON I can actually reason over. That work is what convinced me the patient-directed pattern was real before I started building OpenKP.

This is what patient empowerment looks like in 2026. Not asking for access. Building it.

OpenEP and the Hugotronic Dashboard

I’ve worn an implantable cardioverter-defibrillator since November 2007. The leads (Medtronic 5076 atrial and 6947 Sprint Quattro Secure RV/SVC) have stayed in place across two generator replacements: Medtronic Virtuoso, then Evera XT, and now a Boston Scientific Momentum EL. The simple insight behind the project: leads outlast generators. What matters over time is the health of the leads themselves.

Institutional tools don’t show that view. CareLink, LATITUDE, Paceart, Epic. Each was built for the institution that owns it, in the layout the clinician’s workflow expects. None of them give me an apples-to-apples view of my lead health across vendor changes. So I built one.

OpenEP ingests every interrogation report I’ve collected since 2007 and renders a single-page dashboard at hugotronic.com. Eighty-nine interrogations across three Medtronic and Boston Scientific generators, plotted continuously, with capture thresholds, impedances, and sensed amplitudes against typical ranges and clinical investigation lines. The deploy is a static HTML file served by Cloudflare. No internet connection required to read it.

This is the show-don't-tell version of the case Liz Salmi and I made in our 2025 NAM commentary. The framework is the argument. The dashboard is the proof.

OpenEP is forkable. If you have an ICD or pacemaker and access to your own interrogation PDFs, the pattern generalizes. The harder problem is usually getting your own PDFs out of the institution that holds them. Worth the fight.

AI Patients Community

AI Patients had its first community meeting on March 13, 2026. Fifteen of us. Patients and caregivers from across the U.S. and Portugal, living with genetic heart disease, type 1 and type 2 diabetes, multiple cancers, lupus, cystic fibrosis, long COVID, chronic Lyme. Some of the most experienced patient advocates in the country were on that call.

The meeting included a live demo of Health Skillz analyzing a complex multi-system case in my dad’s record. The AI generated a hypothesis that no provider across three health systems had ever formally evaluated. That moment, on a Friday afternoon, with fifteen people watching, was the clearest argument for patient-directed AI I’ve ever made.

The second meeting is overdue. I’m aiming for late May or early June.

Operationalizing Critical AI Health Literacy

Both of our 2025 papers are now published:

My main intellectual focus for 2026 is operationalizing the framework. Less theorizing about what patient-directed AI could be, more building and documenting what it actually is. OpenKP is part of that. The AI Patients community is part of that. The talk I'm giving at AcademyHealth in Seattle on May 30 is part of that.

If you’re at an organization, foundation, or research group thinking about this, I’d like to hear from you.

Tools I’m Using

  • Claude Code CLI for OpenKP and most serious build work. The shift from chat to CLI agent has been the biggest single productivity change I’ve had in a decade.

  • Claude Cowork for everything that lives outside a code repo. Project planning, document drafting, reading my dad’s records, prepping the Seattle talk. The local-files-on-my-Mac model is the right shape for patient work.

  • Claude for Chrome for browser-side tasks where the data is already on the page.

  • Gemini, NotebookLM, ChatGPT for cross-checking and research. Triangulation is a real safeguard against AI hallucination, and it came up explicitly at the AI Patients meeting as a community practice.

  • Wispr Flow for voice input in English and Portuguese. Most of what I write starts as speech.

Ideas I’m Working With

  • The gap between data access and data comprehension. Federal rules made the data accessible. They didn’t make it understandable. AI is the bridge.

  • Community as infrastructure. The most interesting health AI work right now is happening between patients, not inside health systems.

  • Local-first AI. The data never leaves your machine. The credentials are yours. The agent answers to you.

  • The speed of desperation. As my friend and fellow advocate Sue Sheridan puts it: “Doctors adopt AI at the speed of trust. Patients adopt it at the speed of desperation.” Patients aren’t waiting for institutional permission. They’re already here.

If any of this resonates, I’d love to hear from you.

Let's Work Together