The Patient Who Stopped Waiting and Built It Himself

For more than a decade, Hugo Campos was one of the most recognized voices in the patient-data movement. He fought for the right to read the data coming off the implantable defibrillator in his own chest. He gave talks. He escalated. At one point he and Josh Mandel opened Chrome’s developer tools, diagnosed exactly why Kaiser’s system wouldn’t release his records, and handed leadership a report naming the one-line fix. Nothing happened.

“Why would they bother fixing something that one of 13 million members uses?” Hugo asks. It isn’t bitterness. It’s a clear-eyed read on incentives.

From advocacy to agency

So Hugo did something that surprised us. On the show, Leon recalls what Hugo told them in their pre-call: “I’m not trying to change the system anymore.” On air, Hugo explains why. “I think we have to rethink incentives and realign incentives,” he says. The system changes when the incentives do. Rather than wait for that, he turned to a new generation of agentic AI tools and built what the institution wouldn’t.

The result is OpenKP, an open-source tool that pulls his own records out of Kaiser, the same data the patient portal won’t surface. The remarkable part is who built it. Hugo describes himself as a non-coder. He handed Claude a problem late one night, convinced nothing real could happen in the half hour Claude promised. Two weeks later he had a working prototype reaching his medication list. “I do not know what I built,” he says cheerfully. “It’s a black box, built by Claude Code and verified by Codex.”

Two AIs, checking each other

That verification step is worth copying. Before open-sourcing anything, Hugo ran OpenAI Codex against Claude’s work to hunt for gaps and scrub any PHI. Codex flagged issues, Claude fixed them, Codex re-checked. Leon called it “good healthy paranoia,” and it’s a transferable practice for anyone shipping software they can’t fully read.

Institutional AI vs. Patient-Directed AI

Underneath the build is the idea that gives the episode its spine. Hugo splits the world into institutional AI, deployed by organizations for compliance and throughput, and patient-directed AI, chosen and steered by the patient. Patient-facing is not the same as patient-directed: a slick portal chatbot can still be doing the institution’s bidding.

He is careful about the hype. “It’s not an answer machine, it’s a thinking partner,” he says, insisting that users keep their own critical judgment. In his new National Academy of Medicine commentary with Liz Salmi, he names the skill this requires: critical AI health literacy, which he argues now matters more than prompt-writing.

The honest tension we didn’t resolve: if anyone can build patient-directed AI, what happens to the patients who can’t? Hugo’s answer is that the tools are finally good enough to guide a “pretty lame carpenter” along. We’re not sure that fully closes the gap. But his story is the most concrete argument we’ve heard that the gap is worth closing.

Listen to the full conversation: https://practicalaiinhealthcare.com/episodes/#S1E41

You Might Also Enjoy

• S1E30 with Amy Price — Another deep look at patient advocacy and participatory medicine, and how AI governance shapes who actually benefits.
• S1E37 with Danny van Leeuwen — A patient-and-caregiver activist on steering your own care through the system, a natural companion to Hugo’s agency-first message.
• S1E35 with Barry Chaiken — A physician who became a patient, on what using AI from the patient’s side of the bed actually feels like.