AI-Powered Practice: Scaling Specialty Medicine on 3 Trillion Data Points with Databricks

Summary

Modernizing Medicine built ModMed Scribe, a specialty-specific clinical AI documentation system trained on half a billion patient encounters, achieving 2-second documentation latency and reaching 870 practices, 3,000 providers, and 1 million conversations in production.
The end-to-end architecture on Databricks combines ambient listening transcription, retrieval-augmented generation, fine-tuned specialty models, and AI agents, processing 5 petabytes of monthly data across 900 million patient encounters and 3 trillion data points.
The AI-powered practice vision extends beyond clinical documentation into patient engagement and financial workflows—including billing and revenue cycle—targeting the 28 hours per week physicians currently spend on administrative tasks rather than patient care.

AI-Powered Practice: Scaling Specialty Medicine on 3 Trillion Data Points with Databricks

Physician burnout is driven by administrative toil. A typical clinician spends 28 hours per week on administrative tasks, leaving only 27% of their time for patient care. Modernizing Medicine addresses this with an AI-powered practice vision: intelligent specialty-specific AI assistants that automate clinical documentation, patient engagement, billing workflows, and scheduling. Built on 5 petabytes of monthly data, 900 million patient encounters, and 3 trillion data points, ModMed Scribe combines ambient listening transcription, retrieval-augmented generation, fine-tuned specialty models, and AI agents orchestrated through Databricks' unified data platform.
This keynote covers the end-to-end architecture: how ModMed moved from generic AI scribes to specialty-aware clinical documentation systems, trained on half a billion patient encounters with specialty-specific medical content and terminology. Discover how Databricks Model Serving, Delta Sharing, vector search, and custom model training enable 2-second latency clinical generation, reducing provider documentation time while keeping physicians in control. Learn how AI-powered practice transforms medical workflows across clinical (ModMed Scribe), patient (engagement and reactivation), and financial (revenue cycle) domains.

Chapters

FAQs

What is ModMed Scribe and how does it differ from generic medical AI scribes?

ModMed Scribe is a specialty-specific clinical AI documentation system built by Modernizing Medicine that is trained on half a billion patient encounters with specialty-specific medical content and terminology. Unlike generic AI scribes that rely on general medical language models, ModMed Scribe understands the specific workflows, clinical terms, and documentation patterns of each specialty, producing more accurate and usable clinical notes.

How does Modernizing Medicine use Databricks for clinical AI at scale?

Modernizing Medicine processes 5 petabytes of data monthly on Databricks across 900 million patient encounters and 3 trillion data points, using Model Serving for low-latency inference, Delta Sharing for data distribution, vector search for RAG pipelines, and custom model training for specialty-fine-tuned models. This infrastructure enables the 2-second documentation generation latency required for real-time clinical use.

What is the physician burnout problem that ModMed's AI addresses?

A typical clinician spends 28 hours per week on administrative tasks, leaving only 27% of their time for direct patient care. Modernizing Medicine's AI-powered practice vision addresses this by automating clinical documentation through ambient listening transcription, as well as patient engagement workflows and financial processes including billing and revenue cycle management.

What does the AI-powered practice vision include beyond clinical documentation?

The AI-powered practice vision spans three domains: clinical workflows covered by ModMed Scribe for ambient transcription and structured documentation, patient workflows including engagement and reactivation campaigns, and financial workflows covering billing and revenue cycle management. These AI agents are orchestrated through the Databricks Data and AI platform, providing an end-to-end architecture across the full specialty medical practice.

Full transcript

[00:08] Good morning, everyone. Um I have to say at the start that it's a very tough act to follow after that exciting keynote with, you know, a bunch of amazing innovations from Databricks. But, I'm here to share our journey of Modernizing Medicine and Databricks
[00:24] together in the journey of an AI-powered practice vision that Modernizing Medicine has been on. So, I'll talk about the company a little bit, our journey with Databricks a little bit, and our latest AI-powered practice vision a little bit. So, let's get started.
[00:39] Uh I'm Mike Jayaraman, uh CTO for Modernizing Medicine. So, let's talk about Modernizing Medicine a little bit. Modernizing Medicine is short for Modernizing Medicine. And Modernizing Medicine was founded in 2010 by co-founder of Blackboard, if you know Blackboard, uh
[00:55] Dan Cane, and a practicing dermatologist, Dr. Michael Sherling. And by the way, Dan Cane is right here with us in this room, you know, with all of us here. And uh our mission is to place doctors and patients at the center of care through a
[01:12] very intelligent specialty-specific platform. So, we serve various specialties. Uh our vision is to increase medical practice success and improve patient outcomes. And to that extent, we have a complete 360° technology solution for specialty
[01:29] practices that we provide that we'll talk about today. We are headquartered in Boca Raton, Florida, and we are a global company uh in four continents, and uh our our total employee strength is more than 2,500 at this point in time.
[01:50] So, we are built by doctors for doctors. And what do I What do we mean by that? Right? So, we hire physicians and actually teach them how to code. Right? So, we have like, you know, more than 20 specialty doctors as part of our staff who actually code the medical content for each specialty. And guess what? We have close to like a
[02:06] million lines of code, medical content for each specialty, powered by a very proprietary architecture, which is a very powerful thing to do. So, we are one of the unique EHR and practice management solutions built for the specialty nuances. So, we support the native workflows for
[02:22] each specialty. So, they are it's a very clinically aware, you know, intelligence built for each specialty. And all the AI models that we train, they are trained, you know, on the specific specialty context, the medical content in the specialty
[02:37] context, so that it can understand the specific terminology related to that specialty. So, it's not a generic medical language, right? You have to understand the specific terminology of that particular specialty. Talking about our overall suite of solutions a little bit.
[02:54] We have over 750 million patient encounters captured in our EHR today. By the way, I'm being told as of last Friday, this is close to 900 million. Right? And growing by the day. We support 50,000 providers.
[03:09] You know, again, increasing by the day, as of today, across all specialties, by the way. We process about 5 billion charges, you know, from our revenue cycle management team on an annual basis. And we process about 81 million patient messages in our patient engagement
[03:25] platform that we have. So, that's the scale we're talking about in the last 16 years that we have built. Now, many of you, most of you in this room may know this, right? The providers are facing enormous systemic pressures today.
[03:41] On one hand, they have the significant documentation complexity, and on the other hand, the operational and the analytical needs of the practices are evolving by the day. The market is consolidating and there is a shift from, you know, towards a value-based care system.
[03:59] On the other side, the patient's demands are not decreasing, right? They still expect the same high quality of care with a proper level of communication, you know, high levels of engagement from the doctor's office with with a great technology. The regulatory demands are not
[04:14] decreasing, right? They're not easing up at all. In fact, they're increasing by the day. And our interoperability challenges, for those of you who are in health care, they're also multiplying by the day. So, overall, what does it mean, right?
[04:29] What this means is as per one of the polls and surveys, and by the way, there are many, many surveys out there, a clinician, a typical physician, spends nearly 28 hours a week doing administrative tasks. Which means 73% of a physician's time in
[04:46] a week is not spent with patients. That's a very staggering number. Come to think about it, right? So, the burden is real, right? We are in a very We are caught in a very negative feedback loop in health
[05:01] care landscape. Right? So, the administrative burden, the staff burnout, and the margin, you know, margins are compressing. So, this is completely ripe for an AI revolution. Let's just say that. This industry is completely ripe for an AI
[05:17] revolution. Some of these numbers, you know, are pretty staggering. So, just I just kind of put it out there. Now, obviously, there is a massive shift happening with AI. In January 2025, for the first time in the history,
[05:32] the medical group leaders ranked AI as the number one technology priority. And for those of you who are very familiar with this industry, until that point, EHR usability was the prime factor in an NPS score with practices. But in January 2025, the
[05:49] landscape shifted towards AI being the prime, you know, technology priority. So, how do we achieve this? Right? Let's get into the journey of Databricks and Madaket together. So, we have been partners, you know, we started our
[06:05] journey back in 2018. When we kind of started migrating some of our, you know, Hadoop workloads and the EMR pipelines, you know, the foundational workloads into Databricks. But today, it's a completely a modern AI-powered platform.
[06:21] Right? So, we power our analytics through this, we power our AI models through this particular data platform, and we also have many aspects of data analytics, you know, going on through this platform. This particular infrastructure that we
[06:36] have built has been very critical to the launch of what I'm going to talk about very shortly, the Madaket Scribe, which is the ambient listening AI-based ambient listening transcription product that we just launched. So, bear with me a little bit. So, we'll talk about that.
[06:53] We continue to innovate with Databricks on many, many fronts. For example, Delta Sharing, right? As we speak, we have a product going on on Delta Sharing. It's going live, and that speaks to the power of what Databricks and Madaket has been able to create and moving towards an
[07:10] AI-native healthcare ecosystem. This slide shows our scale of the data platform infrastructure. So, we process about five over five
[07:26] petabytes of data per month. We have nearly 1.7 million data tables, right? Remember, over five petabytes of data means like trillions and trillions of records passing through our ingestion, transform, and delivery pipelines, right, on a monthly basis. We
[07:42] have over 1.7 million data tables, you know, in in our entire Databricks platform with about, you know, tracking of 26 million lineage events. So, this is for the the the traceability and for the auditability purposes. Overall, all of this amounts to nearly 5
[07:59] trillion data points. So, that's that's a staggering number. Our orchestration layer runs about 137,000 jobs on a monthly basis with about 4 and 1/2 million queries that are being executed. All of this runs on a compute cluster,
[08:16] like 120,000 plus compute clusters, which kind of amounts to 143 million hours of cluster computing processing time. Overall, what I would say is that this journey with Databricks has led to our modern, automated, and governed
[08:31] lakehouse that we have created together. And this is going to be a power to reckon with in our AI journey. So, talking about ModMed scribe. So, ModMed scribe is our AI solution for the
[08:47] clinical workflow of the providers. Right? So, what we kind of do in this is So, keep in mind there are lots and lots of AI scribe solutions out there. Right? But they're overwhelmingly generic because they do not understand the specialty-specific nuances.
[09:02] So, that's where there was there was a great opportunity. By the way, Dan Kane is was our lead AI developer on this particular solution and he's a very hands-on leader, entrepreneurial leader. So, what did we do in this, right? We kind of a ModMed scribe uses a a combination
[09:19] of rags and fine-tuned structured data models. What do we What do we do? We kind of transcribe at the end end point at the edge on the actual device through a transcription model. After processing that, you know, the transcription, it goes through a set of
[09:35] models which are orchestrated through an agent and we generate all the clinical documentation from that particular transcription. And by the way, this is trained trained on the medical content that we have. This is where the power of structured data comes in. So, we extract the clinical information and we also
[09:51] generate some free form text summaries. After all this, by the way, we used uh you know, the Mosaic AI custom uh custom model serving and the foundational model serving as well as the vector search as part of this entire technology platform. Right? After
[10:07] processing the clinical information from the transcripts, it kind of passes to the physicians and the physicians can accept or reject or make any modifications with a few clicks and off you go. Essentially, what Man Met Scribe does is it gives their face time
[10:23] between a doctor and a patient's back and you don't have to be continuously typing on an EHR platform or an EMR platform as you might have seen in many of the doctor visits that you've gone to.
[10:39] So, obviously, this launch of Man Met Scribe did not come without challenges. Right? So, we wanted to make sure that um we wanted to make sure that we are able to kind of uh quickly really realize the benefit through the the lineage tracking
[10:54] and everything. So, let me go through this, right? We want to make Databricks what they did is they added this constraint decoding in in on their side. And which basically made sure that we get the output in the correct format. That's number one. The second, they added an identifier to every part of the
[11:11] call, you know, that goes through their particular infrastructure so that we can do the lineage and the flow of the agents end to end. Last but not the least, you know, sometimes these transcriptions and the overall processing can take minutes. Right? In the world of physicians, we
[11:26] don't have minutes, right? It's pretty typically in seconds. So, what Databricks team very adeptly did is they they kind of analyzed the the bottlenecks in the underlying infrastructure and worked with us and we were able to reduce the overall latency by upwards of 50% which was a remarkable
[11:44] remarkable achievement, right? Today, we are in the 2.0 version of ModMed Scribe. We started with 1.0, but we are in 2.0 version of ModMed Scribe and I have to say this at this point, kudos to the entire account management team, the technology team, Yatish and others. You
[12:00] know, they have been awesome in this journey of building this platform together. Let's talk about Let's talk about ModMed Scribe a little bit in terms of what it is trained on. It is trained on more than 500 million
[12:15] patient encounters. Remember I talked about 750 million plus 900 million patient encounters in an EHR. So, it's trained on upwards of half a billion patient encounters. Today, it is live in 870 plus practices with more than 3,000 providers live
[12:33] using it and more than 1 million conversations have gone through ModMed Scribe as of right now. And by the way, the the the response and and the excitement on this particular part of the innovation is amazing to see
[12:48] and we are uh growing by the day in terms of the way it's impacting the lives of physicians and the doctors. So, what does this mean in terms of AI-powered practice? You saw the title on on the the session, AI-powered
[13:05] practice with 3 trillion data points. So, now the journey that was AI Scribe, MM Scribe, now the journey of AI-powered practice continues, right? And what does this mean? AI-powered practice brings AI to every corner of your practice in a very
[13:20] easy-to-use manner. So, for example, when you walk in as a patient in a doctor's office, that's the patient persona, which is the patient engagement platform. Through our patient engagement platform, what we are trying to do is to automate the scheduling pieces of it, the messaging pieces of
[13:36] it, and the reactivation campaigns to reduce the administrative work. Right? On the clinical side, you saw the M M scribe, but we are also trying to do the visit summarization as part of the clinical workflow. On the financial side, what's the biggest thing, right? Accelerating the
[13:52] collections, the revenue cycle management. So, we are trying to accelerate the revenue cycle management through our AI innovation. All of this is powered by an end-to-end analytics platform, which provides insights to help us run these practices and the
[14:09] operations of the practices in an efficient manner. And that end-to-end vision is the AI-powered practice. And how does this work? You know, there is an R&D transformation that has been going on for many years, like for the last few years. At the core of it,
[14:26] there is a cloud transformation. By the way, we were born in the cloud company. So, born in the cloud company, born with a mobile interface, a very touch-sensitive, with a structured data behind the scenes, right? So, all the powerful combinations that you can imagine in 2010. And now the same
[14:41] structured data becomes our moat with respect to going towards AI-powered practice. So, essentially, we were on a journey of cloud transformation at scale. You know, turning into modular cloud-native architecture, powering the multi-specialty platform. We had the data platform, we had the nightly
[14:57] pipelines, ETL pipelines, but we kind of modernized it to the lakehouse and the lake base that we have today. And we also have a de-identified data set. Remember PHI, right? So, we are in healthcare. So, we have to have the de-identified data sets
[15:13] at scale, and that particular de-identified data set is also powered by the same half a billion patient encounters. All of this powers the data lake house. And which is the basis behind the AI layer, the gen AI layer, the agents and
[15:28] the models. And there comes to the fruition the entire AI powered practice. Let's see a short video that will kind of illustrate the overall vision. Meet your doctor's new AI assistant.
[15:44] Actually, assistants. Hi there. At ModMed, we're building the AI powered practice around modern medical workflows with expert AI assistants that streamline what slows teams down the most.
[15:59] They understand complex requests and automatically execute multi-step tasks like patient scheduling and communication, eligibility processing, prior authorization, and claim denial appeals. Some AI assistants, like ModMed Scribe,
[16:15] can translate natural conversation into clinical action so physicians can focus on the patients in front of them. Other AI assistants understand payer policies, handle high volumes of work at once, and promptly return a summarized actionable result that everyone can
[16:32] trust. After practice staff approve an assistant's recommendation, the AI assistant takes care of the rest, coordinating tasks, highlighting what still needs attention, and surfacing any additional recommendations for final approval. Since every action is grounded and
[16:49] cited, humans are always in control. Meet your doctor's new AI assistants from ModMed. Always working, always ready.
[17:07] So, there you go. That's your AI powered practice end-to-end for a specialty practice. So, the AI powered practice starts here. Right? Because ModMed is built for specialists, we cater our AI design caters to the nuances of the specialty
[17:22] practices. Right? How does a doctor How does a practicing physician document? How does a staff support a visit? How do patients, you know, communicate with the practices? How do we process a claim and how do we manage a claim? How do we kind
[17:38] of run a practice? How do the leaders manage the financial performance of a practice? All of this is the essence of the entire AI-powered practice of a specialty physician, right? And the outcome is a very practical one.
[17:54] Save time, save efficiency, right? Give time back to the doctors, give them more face interactive time. Um reduce the financial burden, the financial cost, improve the collections, and all in all, make the lives of doctors and patients
[18:09] better. And who is behind all this innovation is our phenomenal R&D Right? We have a global R&D team, very high growth, high speed, very passionate. And I'm very proud to talk about the culture of this company that Dan Cane and Michael Sherling have created from
[18:25] the get-go. It's a award-winning R&D team. And we all have one common purpose, which is to build what's next in healthcare and to make it possible and to make it matter. Right?
[18:41] And this what Mod you know, what Data Bricks does best is to bring this people strategy, the talent strategy along with the data strategy, and brings it to the awesome outcome of AI-powered practice. Thank you all.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.