Discover more from Pear Healthcare Playbook
Lessons from Alexandre Lebrun, CEO and Co-Founder of Nabla, on building a defensible and safe Language Learning Model in healthcare
Alexandre Lebrun, CEO and Cofounder of Nabla, a personalized AI assistant for doctors
Welcome back to the Pear Healthcare Playbook! Every week, we’ll be getting to know trailblazing healthcare leaders and dive into building a digital health business from 0 to 1.
Today, we’re excited to get to know Alexandre Lebrun, CEO and Cofounder of Nabla, a personalized AI assistant for doctors.
Before founding Nabla, Alex was the Head of Engineering at Facebook AI research for almost four years, and is a seasoned entrepreneur, having sold his previous companies Wit.ai, an AI platform that makes it easy to build apps that understand natural human language, to Facebook in 2015 and VirtuOz, a customer service chatbot, to Nuance Communications in 2012.
In this episode, Alexandre shares his experiences after 22 years and founding 3 startups, why proprietary data is more important than ever for defensibility, tactics on GTM and customer focus specific to healthcare buyers, and opportunities and safety for generative AI in healthcare.
If you prefer to listen here is the recording!
Alex’s journey as a serial founder:
With his dad being one of the first HP employees in Europe, Alex shares that he wanted to be an entrepreneur since age six.
“22 years ago, I fell in love with chatbots.”
Fresh out of engineering school and a lover of language, Alex found simple chatbots online and was hooked on the concept. 22 years ago, while still living in Europe, he founded his first startup called VirtuOz doing customer service chatbots.
Alex calls the product “very broken, but still helpful in a customer service context”.
He raised money for it and moved to Silicon Valley, developing it until it was sold to Nuance Communications in 2012. Nuance Communications acquired VirtuOz when they decided to expand into customer service, developing products at the top of the voice recognition layer they’re famous for. Alex shares that after an entire lifetime dedicated to conversational AI, the more he’s discovered that it’s harder than he thought.
In 2013, two years after Siri was announced by Apple, Alex noticed more and more people trying to build natural conversational interfaces. Those are hard to build when you’re not an expert— and expensive. This was the impetus for him to build Wit.Ai, an API that takes natural language as input and returns structured data for developers to build anything with.
Wit.Ai went on to be acquired by Facebook in 2015, which Alex shares was “the beginning of the big AI wave”. The team was small but filled with the best researchers in the world— a big driver for Alex to choose Facebook as the next phase of his career.
Alex’s time as Head of Engineering at FAIR (Facebook’s Artificial Intelligence Research):
Alex worked directly under Yann LeCun, formerly Director of AI Research now VP of and Chief AI Scientist at Meta, and the team was clearly prioritized within Meta given the seating arrangement close to Mark Zuckerberg and also the independence the team was given to explore research.
“The team that is closer to Mark is the priority of the company; it would change every year or so. Back then, all the desks around Zuck were the AI teams. It was a big priority. We had unlimited resources… It was small, but growing very, very fast and with a lot of ambition.”
There are two AI teams at Facebook. Alex shares that FAIR (Facebook AI Research) has always been an academic research group completely unrelated from Facebook’s product side. The team was measured like an academic research lab - publications and citations. It’s incredible how much amazing talent and entrepreneurs came from the FAIR alumni base.
AI research is very empirical, heavily based on large scale experiments. It involves a lot of computing and software programming, which is the function of engineers to support the researchers. Alex’s role was to help these researchers be successful.
One project he worked on was called Tamagobot, an early version of LLMs. Alex shares that this was one of the first chatbots that was not driven by rule-based decision trees and could handle unsupervised learning with a huge quantity of data.
The second team, called Facebook Applied Machine Learning, is responsible for AI inside Facebook products like ads, moderation, safety, and more features that have been driven by ML for 10 years.
After four years at Facebook, Alex was ready to go back to the product side. They’d done incredible things in research, but there wasn’t a real world application. He started thinking about what industry and what pain points to apply their research to, settling on transportation and healthcare as areas with meaningful problems where you can have a big impact. He talked to hundreds of doctors, visited hospitals, listened to medical consultations…
“And this is how we decided to start Nabla and try to fix some parts of healthcare with machine learning.”
Founding Nabla, Alex’s third startup:
“A third company is completely different from the first one. The first one, you're struggling for everything. Nobody wants to give you money. Nobody wants to work with you. A third company after two exists, it’s different. You make new mistakes, there are new risks.”
Alex introduces a problem he calls the Kim Jong Un entourage trap: if everyone agrees with what you say, everybody wants to invest and nobody will challenge you— which means you can make big mistakes.
On the whole, Alex believes a third startup is much easier than the first, especially to build a team. He’d met smart, hardworking people through two startups and his tenure at Facebook, making it easy to pick from a huge pool of people he trusted. Nabla’s CTO, Martin Raison, worked with Alex at Wit.Ai and at Facebook. Nabla’s COO, Delphine Groll, Alex met through mutual entrepreneur friends and he saw how great the Sheryl Sandberg, COO and Mark Zuckerberg, CEO dynamic worked. He was looking for his Sheryl Sandberg.
Nabla is automating the clerical, administrative, non-medical work in the medical space with a real-time AI assistant working alongside doctors.
As the team looked closer at the healthcare space, the most obvious problem was the physician shortage worldwide. Because doctors undergo 10-15 years of training, it’s not an overnight problem to fix. Alex shares that a huge driver of burnout is the administrative work that doctors shouldn’t have to do: patient records, administrative follow-ups, etc.
The clearest pain point was clinical documentation. Alex shares that on a 10 minute consultation, a doctor will spend 4-5 minutes documenting it at the end. It’s a waste of time and in Alex’s opinion, a loss of quality of care when a computer screen sits between a patient and a doctor. Nabla’s AI assistant looks at what doctors are doing and listens to what the patient says.
The evolution of AI assistants:
Alex explains that first-generation AI solutions to this problem still had humans in the loop. The recommendations might be generated by AI, but then it typically goes abroad where a remote scribe will edit. This whole process takes several hours.
“Anything that is not real time is a completely different product. What we see every day is that physicians want to close the file, close the consultation, close everything in real time…
[If] after dinner, they’re asked to reopen the EHR and finalize the consultation, they hate that. That’s really the problem we’re trying to solve. It needs to be real time.”
Solving the last 20% of the problem:
Real time also enables more value to be delivered because you can do more than just communication.
Alex shares that the space for real time note generation is super noisy right now with dozens and dozens of startups trying to tackle this— but his take is that it’s easy to solve 80% of the problem. Produce a nice note and you’re done. The last 20% is what’s very, very hard— getting the ICD-10 values, the CPT code, the billing, etc. done without mistakes. Precision and safety will be key for a product to survive the market and get real adoption by physicians.
Everything Nabla does is ambient. Alex’s opinion is that the time of physicians dictating into microphones is over. If Nabla gets the ambient audio track of what’s happening, it’s enough to act like a good medical assistant would: figuring out what they should write and what they should do. “Ambient solutions are definitely what bring more value.”
Nabla’s tech team is based in Paris. On the startup ecosystem in France:
Alex moved back to France four years ago when he was still a part of Facebook. 6 years ago, Google had just acquired DeepMind in the UK, and so it was a race to capture European talent. Facebook’s rationale was that European researchers who wouldn’t want to move to the US would want to move to Paris— and it’s worked out well. Facebook’s Paris lab has about 200 scientists, the biggest for Facebook outside of the US.
Alex shares that France had changed significantly in the 9 years he’d been away— Macron is pushing for startups and has made the environment very friendly for startups. There’s a huge pool of talent, AI and other sectors, and it’s much cheaper to hire, making it a great place to build a product.
Tips for early building in healthcare
Don’t start with a product. Start with who will pay, and optimize your company to have that narrow focus.
Healthcare has a complicated GTM strategy with three parties involved: patient, provider, payer. Patients will benefit from the solution, but the payer is paying for it, and if the provider doesn’t agree, nothing happens.
“You’re asked to solve three math problems at once, and doing that as you start without a lot of funding is almost impossible. The good news is— when you make it and find go-to-market fit, then you have a very beautiful life: not so many competitors and good funding.”
Alex criticizes a common teaching from startup schools: start with a problem and build a solution to it. He believes it starts with who is paying. Understanding the business model will help you redefine the problem and therefore find the solution.
Alex admits that almost everyone has made this mistake, including Nabla— the fallacy that if the solution is so great, someone will pay for it.
“As entrepreneurs, we tend to be biased towards patients, which is natural… But sometimes, it’s easier to do something that will be the first use and paid for by providers…
In the longer term, you can have a more scalable business follow where payers would pay for it or patients themselves or employers.”
If Alex could do it again— he would start out by becoming embedded in any medical team and conducting bottom up discovery.
Alex shares that he would want to start out by sitting in on a health delivery setting for a few days, listening and observing. By building a great relationship with the first customer, he believes you’ll find very interesting problems to work on.
Regardless of the industry, Alex believes talking to people is the most important starting point, and it’s an area he’s made mistakes in before. As a founder, the first 12 months should be spent working on product and talking to potential customers— and nothing else.
“It’s easy to be spread thin across many things that actually don’t matter at all. I think the most important thing is to really be obsessed with users and product and absolutely nothing else…
When you start to get a little bit of traction, many people get interested in you. You’re happy because it’s the first time people have an interest in your company. You tend to say yes to everything and start to forget what the most important thing was.”
Avoid distractions— like large incumbent companies who want to integrate you into your product or consulting companies who say they could resell your solution.
In Alex’s opinion, these sorts of conversations often don’t go anywhere. These businesses often show a few young startups with exciting products for a PR purpose— including your product as an add-on, etc. — but at the end of the day, their incentives are their own, not yours. Time is precious for early startups; “as a startup, you cannot wait two years to get your first customers”.
How do you know when partnerships with a hospital / payer’s innovation arm is worth pursuing?
Ask startups that have worked with them before.
Alex shares that it’s hard to get credibility and get started, which means you often have to take the risk of potentially wasting your time for two years because you don’t have a choice.
Alex’s preference is to invest heavily in one potential partnership instead of talking to 5 different ones and not having time to follow up.
When it comes to prioritizing a partnership, his advice is to talk to founders that have worked with them before. See how they connect innovation to business operations and actual deployment, whether or not the innovation team is totally disconnected from the business or embedded within.
Alex’s perspective on AI & healthcare:
The biggest leverage of AI in healthcare is not diagnosis, but the follow up with personalized treatment.
Alex believes that AI should do the 50% of clinicians’ workload that is low value tasks… “and so it should be possible to make any clinician at least twice more efficient with AI.” For now, Nabla is focusing on clerical tasks but will eventually move to clinical tasks.
One mistake many people make is to just focus on diagnostics. From talking to hundreds of physicians, Alex has learned that diagnostics is the “easy part of the job”. The challenge is figuring out treatment, follow up, and consistency beyond a few consultations here and there.
Alex believes AI assistants can be most useful at this stage: not making the core medical decisions, but executing on treatment and making sure the patient is reacting well.
We’re moving towards deeper automation of more tasks with high precision. Companies need strong technical teams and proprietary data sets, and they need to understand what’s working and not working for their users.
It’s easy to do 80% of the work — a nice demo with a public LLM — but Alex believes companies need to fine tune their own elements. Fine tuning an open source large regression model with your own high quality data and your understanding of the business will push healthcare AI products past text notes towards real automations of clinical tasks.
Alex isn’t saying you should build your own LLM. AI healthcare startups have to fine tune these LLMs.
Hold on— what is fine tuning?
Alex uses ChatGPT as an example: if you ask ChatGPT a question, it will output two different answers and a human chooses which one is better. The second stage is using ML techniques like reinforcement learning, where feedback changes the behavior of an LLM, to have the LLM select which is the best output out of thousands.
The second stage of fine tuning is where you can really change the behavior of your LLM.
AI comes with risks. In healthcare, mistakes can be dangerous. Safety should be a big concern.
Using ChatGPT as an example again, even factually incorrect answers can come out in perfect form, making it easy to make mistakes. If your product works 80-90% of the time, it’s not ready for use.
“Being wrong 1% of the time is not possible in healthcare. A specific danger of AI in healthcare is that it’s usually right but not always, but because it’s usually right, you may deploy your product too quickly.”
Alex shares that when GPT-3 came out, Nabla tried many experiments: building a chatbot that works as a physician. It was initially very credible, but after five minutes of a test conversation with a patient, told the patient “you should kill yourself”. It came out of nowhere.
“Even if it happened only one out of 1 million times, you cannot accept that. Generative AI is very dangerous… It's technically deterministic, but actually in practice, it's not really deterministic. Using that for marketing is easy — SEO, nice communications — but using this for our patients, you cannot afford to have even 0.1% mistakes.”
On the landscape of healthcare startups overall:
We’re in a bubble right now. Alex believes it’s extremely easy to raise seed rounds at the moment, but Series A will be much harder when it comes to real engagement, retention and growth with physicians or patients. Especially with generative AI and what OpenAI does, it’s easy to build a prototype that makes it easy to raise a seed round.
“The reality is very different between a digital demo and a product that is used at the hospital every day…. I expect maybe a year from now, half of the startups will be dead.”
Thanks for reading Pear Healthcare Playbook! Subscribe for free to receive new posts and support my work.
A note from our sponsor: PacWest
Looking for guidance, connections, resources, opportunity? Pacific Western Bank’s banking products and services are built to support your evolving needs as you navigate the challenges of growing a successful business. As you continue to scale, our team will be with you every step of the way. Ready to take your business to the next level? Learn more: pacwest.com