Bench, Bedside, Garage
A bench prototype is allowed to lie about being finished. A patient is not.
Every medical device I ever brought near a patient started on a bench, pretending to be finished before it was.
That’s not a confession. That’s how the work gets done. A bench prototype exists to prove a concept can behave the way you think it will. It is not required to survive contact with a real patient, a real nurse’s workflow, or a real hospital’s security review. The actual discipline of device engineering, the part that takes years and the part nobody puts in a demo video, is the distance between “it worked on the bench” and “it’s cleared for clinical use.” I spent most of four decades living in that distance.
This week I built something that drifted toward looking finished before it was, and then I pulled it back.
The Bench
I’d been building an interactive demonstration of AI MedAgent’s reasoning method: an AI layer evaluating a post-CABG ICU patient and recommending a medication change. Partway through, I pushed further and wired a version of it to genuinely query live federal data sources, deployed it, tested it. It worked in pieces and then it didn’t, the way early integration work always does. So I made the call any device engineer makes at that stage: pull it off the bench, be explicit about what’s a concept demonstration and what isn’t, and don’t let anyone downstream mistake a prototype for a finished product. The site now says plainly what’s illustrative and what’s a scripted example. That’s not a retreat. It’s the same rule I’d have insisted on if this were an infusion pump instead of a website.
The Bedside
Here’s the part that actually matters, and it’s bigger than my website.
That gap, between a demo that looks transformative and a system that survives production, is the entire history of health IT for the last twenty-five years. CPOE was supposed to eliminate medication errors. It caught some and introduced alert fatigue that clinicians now click through by reflex. Interoperability initiatives were supposed to let a patient’s data follow them. Data followed the patient into more fields to populate, not fewer decisions for a clinician to make alone. The Surgeon General’s own workforce advisory puts nurse documentation time at close to 40 percent of a shift, not because nurses got slower, but because the systems that were supposed to save time moved the paperwork instead of removing it. None of this was fake. All of it worked on the bench. Most of it never closed the distance to the bedside, and the industry kept shipping anyway, because shipping the demo felt like progress, and admitting the gap didn’t.
AI is about to do this at a scale and a speed nothing before it has matched, because nothing before it has ever been this easy to make look finished. A model that reasons fluently about a clinical scenario is, on a bench, indistinguishable from one doing it for real. The distance to the bedside hasn’t gotten shorter. It’s gotten harder to see.
The Math
Here’s why that distance matters more now than it ever has.
By 2030, for the first time in this country’s history, there will be more Americans over 65 than there are children. By 2050, nearly one in four of us will be 65 or older, up from 17 percent today. That’s not a forecast anyone’s arguing about. It’s already baked into who’s alive right now and getting older on schedule.
Per-person healthcare spending on someone 65 and older already runs more than five times what we spend on a child, and over double what we spend on a working-age adult, by CMS’s own numbers. National health spending crossed $5.7 trillion in 2025, up 7.3 percent in a single year, and federal actuaries project the share of the economy it consumes will climb from 18 percent today past 20 percent within the decade. Spending is growing faster than the economy that has to fund it. Layer a population that’s aging into the group that costs the most onto a spending curve already outrunning GDP, and you don’t get a slow decline. You get a curve that bends toward collapse, on a schedule that’s already written.
You cannot staff, document, or bill your way out of that math. The only lever big enough to matter is spending the same dollar and getting more outcome for it, which means the reasoning has to get smarter before the patient gets sicker, not after. That’s the whole premise of everything I write here. It’s also, not coincidentally, why I keep building demonstrations of a reasoning layer instead of another dashboard.
The Garage
Am I the person who fixes this. No. I’m not twenty-five with nothing to lose and a garage. What forty years of watching this exact failure mode, demo outruns deployment, promise outruns proof, across every wave of health IT connectivity I’ve touched, closed-loop infusion integration included, actually buys you isn’t the ability to build the fix alone. It’s the ability to recognize the pattern fast, including, this week, in my own work, before anyone downstream had a reason to trust something that wasn’t ready yet.
But recognizing a pattern isn’t the same as accepting it. This is fixable. Not by patching what’s already there, another dashboard, another point solution wired onto the same workflow, the kind of incremental thinking that’s given us a $5.7 trillion system with outcomes that haven’t moved in step. Fixing this takes the other kind of thinking. Vision. A real appetite for risk. The energy to look at something everyone agrees is broken and refuse the easy version of solving it.
AI MedAgent, at the concept stage it’s honestly at right now, is my attempt at that different lens: a reasoning workflow, not a chat window, built to weigh outcome against cost the way a good clinician does, faster than any clinician has time for. Whether it or something like it turns out to be the actual answer, I can’t promise you. I can tell you it’s the right shape of question.
Which raises the one I don’t have an answer to. Are we putting our money where that kind of thinking happens? Healthcare has never been friendly to a twenty-five-year-old with nothing to lose and a garage, too regulated, too slow, too expensive to fail fast in. But reasoning that’s cheap and capable enough to sit inside a clinical workflow instead of a chat window is genuinely new. It wasn’t available the last time anyone tried to fix this. Whether that’s enough to break twenty-five years of pattern, I don’t know. Nobody does, not really, not until somebody with a different lens than mine, in a garage or a lab or a resident’s call room somewhere, decides this is the problem worth staking everything on.
I still believe someone fixes this. I just believe, forty years in, it’s more likely to be someone who hasn’t spent forty years being told how healthcare works.