How I Ended Up Here

tl;dr

Decade-and-a-half from pop punk guitarist / bad MMA fighter / pizza delivery guy to Head of Clinical AI at Suki. By way of: computational biology grad school, med school, residency during the pandemic, an incomplete UCSF informatics fellowship, a failed founder run that Epic crushed, a brief HealthcareAgents detour, and a year-plus building out Clinical AI at Ambience. People keep asking how I got here. The honest answer is mostly luck and being early on a few things. Don't do it the way I did it lol. But if any of this helps someone trying to break into Clinical AI, that's the point.

Punk, Pizza, MMA, and a Senior Thesis

When I was an undergrad I had no idea what I was doing. Honestly didn't want to be a student at all. I was more focused on playing music in a semi-successful pop punk band and training MMA, and that is what I thought my career was going to be. My parents begged me not to drop out. Then in my senior year my MMA coach went through some tragic stuff in his personal life and pulled me aside for a heart to heart, where he literally said "Martial arts are not lucrative, don't quit your day job, Karl." 😅

I didn't really know what to do at that point but I was getting pretty interested in molecular and computational biology. I wrote and presented my senior thesis on a tumorigenesis model heavily inspired by Michaelis-Menten kinetics. It was a terrible paper externally, but internally it made sense to me. After I presented it at Senior Seminar, I ended up walking out with the chair of the biology department, who I mostly got along with because I was an ex-wrestler. He said something like "I am not sure if that is genius or complete nonsense." Hard to tell tbh, because the paper was entirely conjecture strung together from bad formal logic, bad math, and a literature review I had put a lot of effort into.

Either way, out of pity or curiosity, he asked what my plans were after I graduated. I had none. He told me I could apply to grad school and it would be a funded position with a small stipend. My day job at the time was pizza delivery guy, so that seemed strictly better. Plus my band had just broken up after getting dropped from a tour we were supposed to go on in Japan with I Call Fives, The Wonder Years, and Four Years Strong (name dropping these for my pop punk / scene homies).

Quick application, got accepted.

Two PIs and a Dying Dog

Got assigned to a PI who I won't name. I think he took some pleasure in the fact that people knew him as a mean guy. To be fair, he was very sardonic and quite bright, and I was definitely not the star graduate student. So I get it. He still made my life a living hell while he taught me good lab practices and, serendipitously, made me learn Python to do my own data analysis (computational and integrative biology program after all).

I was still pretty focused on MMA at the same time. I remember one Monday after I was supposed to have spent the weekend learning some Python: I had been training instead. He asked how my Python was going. I said great. He asked me to print hello world. I could not lol.

You get the picture. He was kind of a prick. So was I.

Things eventually came to a head when a bunch of my fusion PCR and RNA experiments were failing. I was a lab klutz for sure, but I was confident I was doing everything right. After a few verbal lashings, 12 to 14 hour days, weekends in the lab, he finally realized the gel recipe he had been giving me was wrong. He had done some "customization" of the recipe to cut costs, which turned out to be the reason for all the failures. I was livid.

Chugged along for a couple more months. Then I had an overnight gene expression experiment where I was sampling embryos at different timepoints to track some microRNA expression we thought might inhibit a gene called lefty or something to that effect. The details don't matter. What matters is that I had slept in the fucking lab on and off in two-hour intervals all night to hit my sampling timepoints. The PI came in the next morning. I had 14 of 16 samples done. My dog was defecating blood at home. I told the PI I needed to go take care of him and asked if he could help me finish the last two samples. He said something like "This is your experiment, what is more important, your career or a dog?"

I quit his lab the next day.

Luckily, after talking with the chair (wrestler bond once again coming through), he said the department had received similar feedback from other lab members and they could just transfer me. A purely computational lab had an open spot, which was great news because one of my best friends was already in that lab and I was honestly kind of bored of wet lab anyway. I really enjoyed Python and data analysis. The new PI was intimidating in a different way. Man of few words, math and programming olympiad. Very hands off. That's how I learned the term and the difference between a micromanager (my last PI) and someone who is almost entirely hands off lol. Great guy overall.

From Insect Models to the Wards

I did a bunch of fun research in that lab and met some of the best people I have ever met in my life. The problem was I was bored of insect and worm models. What got me interested in the first place was cancer research in undergrad, and around that time I watched my grandfather slowly starve to death on a hospice bed from a tumor of unknown origin. "How could you not know the origin?" I was upset for a long time. I wanted to pivot my research from basic fundamental work on non-coding RNA to translational human cancer research, any way I could.

My PI sat me down and was honest with me. I was in the wrong lab for that. If I wanted that path I'd have to do a translational postdoc, probably several, if I ever wanted a tenured academic job one day.

I shit you not, that same week a postdoc who was probably the smartest dude I knew (he taught me what touchdown PCR is off the dome, that was crazy) told me he wanted to quit and open a food truck because he was sick of chasing tenure and making no money. I think less than 10 percent of PhDs at the time got tenure, and industry wasn't viewed as "making it" yet. Tech and biotech weren't really sexy, they were for people who couldn't hack it in academics, or so I thought.

Anyway, back to my PI. His other suggestion was that I could skip all that, go to medical school, and "get rich." I wasn't a stranger to medicine. My dad is an anesthesiologist, and I had rejected becoming a physician for a long time because of a difficult relationship with him growing up. We had mended things by then, and he had been begging me in undergrad to give medicine a shot. I think my PI was the push I needed. On a whim, mid-application cycle, I got my stuff together and applied. By the grace of God (or probabilistic luck, I am an atheist), I got into the school that was literally five minutes from my house.

Going in, I thought I was going to be an oncologist doing medical genetics and genomics research for a career. I had picked up solid CS skills in grad school, took DSA and probability/stats courses, kept up with programming as much as I could. Was admittedly hard though. I matched into my number one ranked program near West Philly. I won't say much about it because it produced some of the worst and simultaneously best years of my life. I will say this much: extremely malignant program, during a pandemic, in the Northeast, refrigerator trucks outside the hospital, BiPAP machines being used as vents, all of it. Terrible time.

To be clear, I don't think I was the best resident ever. But even the actual best residents at that program were treated like animals. There were silver linings though. I met my wife there. I learned to be a halfway decent hospitalist.

The Doctorlingo Detour

Because there was no real opportunity for oncology research during the pandemic, I serendipitously got very interested in NLP. I met a guy on Reddit talking about COVID who turned out to be an EM doc and PI at a Harvard CS lab. He wanted to build a repository of patient-friendly definitions for medical terms. They were crowdsourcing those definitions through a website called doctorlingo. Traffic wasn't bad, around 6k unique visitors per month. The math on the timeline, though, was rough. I mean, by my calculations it would have taken 20+ years to reach a meaningful number of definitions at that rate.

We got to talking. I don't remember who pitched the idea of using the repository to automatically simplify clinical notes, but I do remember being the one to say "hey we can just do that for the definitions themselves." There had to be a way to machine translate biomedical definitions into patient-friendly ones.

This project changed my life. I built a really bad recursive method first. Then I learned about transformers. BERT and T5 were the big models at the time, and I started using those, but a little model called GPT-2 had also just come out. I realized pretty immediately you didn't need the encoder models, and I started finetuning various clones of GPT (GPT-J 6B among them) to perform sentence simplification on biomedical definitions. I was so impassioned by this work, and so disappointed with the state of technology in medicine for both clinicians and patients (the NLP tools I was using were amazing, to be clear; the clinical tools were the problem), that I realized I could probably make a much bigger impact doing an informatics fellowship than oncology.

So that's what I did.

UCSF, 2022

I wound up matching to UCSF. For a dude who did not know graduate students could get paid, who went to a public undergrad and a small DO school down the road and then a community-ish residency, matching to UCSF felt surreal. I had been following Wachter for a long time. I also was not sure I was ready to leave Philly. I had spent the better part of my life there and had strong roots.

Still, the opportunity wasn't one I could pass up. So I sold or threw away basically everything I owned (I had also just gotten married and my wife wouldn't let me drive a U-Haul because she was convinced I would die in it) and moved into a 500 sq ft apartment in Nob Hill that literally bordered the Tenderloin. Sight unseen, by the way. My residency wouldn't let me take off the last rotations to go visit lol. I was on inpatient and ICU for most of the back half of third year.

The year was 2022. There was a sense of joy I could feel in the city as people were getting used to post-lockdown life, and all of that combined with being in SF as a tech hub and being an informatics fellow at a top 3 medical research uni made me unreasonably excited. I was genuinely taken aback that there were signs about data all over the place. There was an old-school theater with the cool black-on-white text signage that read "Data After Dark by Snowflake." I felt like I had arrived. To quote one of my favorite bands that I got to play shows with as a kid: "This is the first time that I've seen exactly where I want to be, and how the fuck I'll make it there."

If you know anything about AI and 2022, you also know things were about to go bananas. I was in the right place at the right time. I was also probably one of like 10 to 50 physicians in the United States who had read the InstructGPT paper earlier that year. I probably sounded insane to a lot of people because I would not stop telling them what was coming. "You will just be able to talk to ML models and they will do what you want soon." People looked at me like I had 5 heads. I remember my PD gently roasting me with something like "Karl is going off about trigonometry again" when I was ranting about cosine similarities lol.

Btw, this PD, after a bad first grad school PI and a malignant residency PD, was honestly the first leader I had where the blend was right. Not a micromanager, still gave solid guidance and mentorship. I really respected him. If you know the PD of the UCSF CI program, you know he is probably one of the best and genuinely most caring dudes you will ever meet. Everyone should get to experience leadership like that at least once.

Things were going great. Rapidly published a paper, co-authored a call to action around Dobbs, set up a bunch of fun CI projects (data analysis on hospital staffing throughputs, helped design an ML framework to identify strong hospitalist dischargers). I got to meet folks like Julia Adler-Milstein, Atul Butte, and Bob Wachter, which if you know anything about the space you know why that is so cool.

Eventually, a serial entrepreneur came after me to start a company. He was the former chair of EP at UCSF and the guy responsible for creating the predecessor to the Watchman device. I was not totally cold on the founder path. I had interviewed for a Meta AI fellowship the year before, though I did not advance past the first technical round. But the academic plan was the plan. I was convinced I was going to be a clinical informatics researcher. So convinced that I had already lined up a T32-style postdoc through UCSF's CRISP program under the tutelage of Atul Butte and Vivek Rudrapatna.

Look, leaving that on the table was probably one of the hardest decisions I have made. I had finally landed in a supportive academic environment. Real shoutout to the CCIB folks back in grad school as the other exception, but from that first grad school PI through med school and residency, the people I had been around in medicine had honestly been awful, pretty consistently. The UCSF year was the first time in a long time that the people part of the job was actually good.

I took the leap anyway. Founded a company with a guy I barely knew, in a space I barely knew (I knew the ML and the medicine, the company-building part not so much).

Quench, and Epic's Big Squeeze

We called it Quench. We were meant to quench physician burnout, or at least that is how it was pitched to me. Building it was fun and stressful, and I learned a lot. We got some cool prototypes shipped and had a pilot site ready to go.

This is also where, beyond my user-level hatred of Epic and after learning about the HITECH Act basically handing Epic the keys to the castle, I developed an actual disdain for the company. Hard lesson: all is fair in love, war, and capitalism (I'm a leftist btw). Epic does not want innovation in healthtech. They want money. They announced they would build an in-basket manager, and our pilot site dropped us immediately. I later learned Epic had no in-basket manager at the time of the announcement, and when I previewed what they ultimately built, I knew it wouldn't work. They crushed a company that might have actually helped with this problem so they could wedge themselves into an AI market that scribes were already dominating.

My co-founder was much more business-savvy than me and convinced me we had to find a faster path to revenue. We landed on a pivot to qualified medical examinations. The tool we had built was a RAG app with agentic features that worked most of the time lol. It could easily be served to medical professionals without needing an EHR integration, and was a much more direct path to revenue.

Deep down I knew I wasn't excited about it. I helped build it anyway. Finished the MVP, got our first set of alpha users, got real feedback. Clinical AI companies were starting to boom at the same time, and my in-basket was getting hit with outreach daily. I felt bad about it, but my heart just wasn't in building tools for what were essentially legal processes. I wanted to build clinical AI at the point of care.

I got an offer to come on at a Series A startup called Encultured AI, who were pivoting into a new company called HealthcareAgents. They were all super talented CS and math people from the Bay's finest institutions and were deeply connected in the ML and AI research space. I still think those are some of the smartest and coolest dudes I have ever met. One of them was previously an early member of technical staff at OpenAI and had laid groundwork for the research that powers most of OpenAI's RL work today. The CTO was a beast of a CS PhD, very insightful. The CEO was a prominent AI safety researcher at Berkeley. There was also a Thiel Fellow on the team. (Not an endorsement of Thiel, who has very clearly lost the plot. The selection bar for that fellowship is brutal though, and this person earned it back before the dark-enlightenment chapter of Thiel's career had really started.) And then there was me, some random clinical dude they found online who knew about RAG (again, being early helped me here. While most of my clinical colleagues were finding out what an LLM was, I had built and shipped a production-grade agentic RAG app in 2023 lol).

Still, I just did not understand their vision for a D2C services play. They were extremely tolerant of me telling them that and were willing to listen to my roadmap and product ideas, but I ultimately got recruited by Ambience to come on as their first full-time physician hire.

Ambience, the Big-Boy Tech Job

Ambience was doing real work at the point of care with doctors and LLMs and had a pretty cool partnership with OpenAI's applied research lab. They had less ML talent if I'm being honest, but their head of engineering was a very thoughtful former MLE at Cruise from before MLE was really a job title. I also knew the part-time physician hire there who was coming on full-time in a couple months, and I really admired that guy. He had less LLM experience than me at the time, but I could just tell he was sharp. I remember telling him most of what I do isn't actually that hard and he'd be better at it than me in three months. I'm pretty sure that turned out to be true. He's a beast. Great command of quantitative and product reasoning. I don't know if it's from studying physics, or that big-brained people just tend to study physics lol. Either way, awesome guy.

For not having many clinical folks (a couple of very talented scribes), I was really impressed with what Ambience was pulling off as a Series B. They were punching way up, winning deals against Abridge in head-to-head vendor pilot evals. What I think they got right early is mostly visible from the outside. The engineering hiring bar was unusually high for a Series B, all senior to staff level ex-FAANG, ex-Palantir, that kind of resume, and you can see it on their LinkedIn page if you scroll their team. The product reflected the bar. The harder thing to see from the outside is the culture. The cross-functional respect across engineering, product, and the clinical team was the highest I have seen anywhere. The level of depth, commitment, and thought that group brought to the work was biblical. Some of my favorite people to this day. Love those guys.

Zooming out, this is the part most healthtech companies get wrong, and it is not actually a tech-specific problem. If you have ever worked in a non-tech environment with a manager who has never spent a day in your shoes telling you how to do your job, you know how insane it is. Building clinical software without giving clinicians a load-bearing seat at the table is the same dynamic. Epic is the canonical example. Their EHR is widely cited as one of the worst-scoring pieces of clinical software on standard usability instruments, with SUS scores in the F range across multiple published studies. The companies who win the next decade of healthtech on the product side will be the ones who treat clinician contribution as a first-class input, not as a focus group held the week before launch.

Alright, enough glaze. Back to me. I already knew a fair amount about LLM evaluations and LLMs in general going into Ambience, but I really sharpened my craft there. Quench had been a very different feel: founder and everything-else-person at a seed-stage company. Ambience felt like my first real big-boy tech job. We grew the Clinical AI team by several physicians, RNs, and PAs over a few months, and won a massive academic center deal beating Abridge and Suki as the clear underdogs. Our product quality was clearly better as well, which bore out across several metrics from the Cleveland Clinic.

I stayed there a little over a year and watched Ambience go from underdog to serious threat in the market. The Clinical AI team grew from a handful of folks to around 20 in that time. I got really solid at interviewing and evaluating for an archetype that has come to be known as "Clinical AI."

Despite all that, I was starting to miss home and the East Coast deeply. I asked my manager (also an awesomely smart dude: UCSF-trained internist with an NLM fellowship in ML at Columbia. I genuinely have no idea how I wound up on that team) if I could go remote. He said probably in a year or so, when they opened the NYC office. I couldn't wait. I had multiple competitors and other healthtech companies in my in-basket daily. My wife had just gotten laid off. We wanted to buy a house, have kids, settle down near family, that whole thing.

So I quit. No real plan. Big ups to my manager for making sure I wasn't destitute by giving me completely undeserved severance.

The Tale of Suki

When I say "no real plan," what I mean is I had not committed to a company. 2025 was the year of the scribe lol, and I was riding the wave. The scribes weren't just scribing anymore. They were doing RCM, building chart chat bots, expanding everywhere. That meant a lot of startups looking for a wedge, and some bigger companies too, wanted a piece of yours truly. It was honestly a great feeling.

But I was pretty confident I was going to one company in particular: Suki. They were a competitor to Ambience. To be honest, I viewed them as 3rd or 4th in terms of product quality, with Ambience #1 and Abridge #2. I knew eventually all of these companies would catch up to each other on product. Suki had two very interesting things going for them. One was a wild approach to capturing niche areas of the market through platform deals. The other was my best friend from fellowship, Stefano Leitner.

I can't really capture Stefano in words tbh. I use the words "insightful" and "thoughtful" a lot when I'm complimenting people. Maybe because I struggle with both unless I am wildly interested in something. Stefano is the most thoughtful guy I know, full stop. I remember when I first met him I did not expect him to be that way either. He's a big hulking dude lol and loves EDM (which I respect, I am also a meathead at heart and love EDM). Anyways, Stefano is the kind of dude who will fly across the world to come watch me get married, after knowing me for just one year, and then hop on a plane immediately afterward to go explore Japan and surprise me with a gift from the official Nintendo store of my favorite game that I had introduced him to. He takes suggestions seriously too. When I mentioned in passing during fellowship that he should play Zelda, I did not realize he was going to try to 100% Tears of the Kingdom like I did and become absolutely engrossed in it. He's one of the best dudes I've ever met, we had worked together on Quench, and I was so excited to get to work with him again. AND BE FULLY REMOTE.

I've now been at Suki for a year. It is very comparable to Ambience. Some of my early opinions on Suki's product being weaker than Ambience's were true. On others, I now think Suki has the advantage. Suki is my 4th job in tech but what I'd consider my second non-founder big-boy IC role, and there are some really interesting differences between it and Ambience.

Ambience was high energy, full of fun people, tight-knit. Also somewhat of a live-breathe-eat-sleep-work and then party with the same work friends mentality. And don't get me wrong, Suki does that too. One of the most epic work parties I've experienced was our annual offsite (tequila and a Bollywood dance party went absolutely crazy). But the culture is also genuinely endearing. People mean well. I think I went from being one of the more type B people at Ambience to one of the more type A people here. People are patient and kind and really focused on the mission. They celebrate wins and lift each other up. There is also a solid distinction between work and life that is encouraged here, which I appreciate now in a way that I don't think I would have a few years ago.

The talent is also solid. Our commercial team is one of the strongest I've ever seen. Our CMO is a business development genius. I got to work on a lot of our note quality efforts, and I can comfortably say we are on par with or better than our competitors on most specialties. We recently pulled ahead in KLAS ratings for the first time. We also took a big swing where not many competitors were willing to go: automatic order placement from the transcript. I have seen this project abandoned at multiple startups. Our first swing went poorly, that's for sure, but with a few iterations, utilization climbed and the ML predictions got better. We are playing some catch-up on patient and chart chat agents, but there are some pretty cool things coming down the pipe for that and for general medical knowledge. RCM is going to be second to none soon (I think we may actually wind up winning a lot of that space). And the platform play remains one of the most distinctive things about this company. We are crushing that space and have virtually no competitors in it.

I'm also pretty excited that a lot of my hard work got recognized, and I recently got promoted to Head of Clinical AI. Moving from an IC role to a management one. That's actually what prompted me to sit down and write all of this.

I have been getting outreach from clinical folks for years now. Even back when I was just doing ML consulting on the side in residency and fellowship, people would ask me "how did you get to where you are?" Now, after a near decade-and-a-half journey since starting grad school (closer to two decades if you count undergrad), I am finally in a position where I actually feel comfortable mentoring people and giving advice.

My path was weird and full of hard luck. It should be smoother for others. I don't recommend you do it the way I did it. But if any of this helps inform or inspire someone trying to break into this growing field of "Clinical AI," that's the point.