I've lived three technology waves in healthcare. Safety didn't appear till regulations forced it.

I was there when electronic health records transformed the industry. I watched the technology get applied in ways that served capital incentives more than patients or clinicians. Billing systems got inflated, and a disproportionate share of technology attention went into revenue cycle management, the cat-and-mouse game between providers and payers. Just because a technology is here does not mean it will be adopted for the greater good.

At Google Cloud, I was the healthcare GTM lead. The job was translating platform capability into market entry points, reading which problems had organizational readiness, not just technical feasibility. That experience gave me a window into how a technology wave reshapes not just healthcare but entire industries. The dynamics of AI deployment rhyme with what I observed during the cloud wave, but with higher stakes and faster timelines.

At Carrum Health, I led an organization-wide AI transformation for the technology and care operations functions. The strategy included voice agents, SMS-based nudging, real-time in-call assistance, and a conversational analytics layer. Four principles emerged: standardize before you automate, choose the right human-AI model (Centaur vs. Cyborg), maintain a skip list of what not to automate, and prove ROI with flat-file pilots before committing engineering. I saw, first-hand, the gap between "this works in a demo" and "this is running reliably in production."

The agentic AI platforms I evaluated at Carrum all seemed to invest in monitoring and observability. They can tell you what the model did. But that observability didn't seem to be made specifically for healthcare contexts. The monitoring frameworks these platforms use were built for general customer service, not for clinical workflows where the consequences of a wrong action are fundamentally different. Getting observability right for healthcare is itself an unsolved problem.

Beyond observability lies a harder question: why did the model do what it did? That is the domain of interpretability, and it is where researchers like Chris Olah, Joshua Batson, and Neel Nanda, and teams at startups like Goodfire.ai, are doing foundational work. I have followed it closely, and I believe it represents the path to genuine reliability and safety in high-stakes domains like health. But interpretability research, like observability, still needs industry context to be actionable. The people doing this science need collaborators who understand the nuances of clinical workflows, patient safety, and the regulatory environment. I see myself as a supportive team member who can help AI researchers bridge the gap between research output and clinical practice.

I trained as a physician. I worked in clinical environments where software is part of the daily routine of care delivery. I have a visceral, not theoretical, sense of what is at stake when a probabilistic system enters that environment. Most people reasoning about AI safety in healthcare do so abstractly. I've been inside the workflow. That is a different kind of context.

Operator path: Working inside or closely adjacent to a frontier lab or company to help make advanced AI systems safer in high-stakes domains. Healthcare is the first such domain where I can contribute uniquely. This path feels more concrete and more closely tied to my existing experience.

Institutional path: Helping build governance structures, standards, and evaluation frameworks for advanced AI risk across sectors. This is a genuine interest that I want to explore further.

Both paths need people who have been in the room where deployment decisions get made.