# Abridge used UPMC funding and data to build its early AI

Markdown copy of https://realalpha.app/s/abridge-used-upmc-funding-and-data-to-build-its-early-ai/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Shiv Rao](https://realalpha.app/builders/shiv-rao/) · Shiv Rao is cofounder and CEO of Abridge; practicing cardiologist. · Co-founded Abridge, which brings context and understanding to medical conversations, and practices as a UPMC cardiologist.
- Host: [Alex Konrad](https://realalpha.app/people/alex-konrad/) · Founder & editor, Upstarts Media · Ran Forbes' Midas List for ten years and co-created its Cloud 100 list; founded Upstarts in 2025.
- Show: Upstarts
- Episode: [Abridge’s Shiv Rao: Taking On Microsoft To Cure Doctor Burnout With AI](https://realalpha.app/v/W4_aRWsHogM/), Jul 23, 2026
- Moment: 12:04 to 17:07 of the episode (5 min), playing on https://realalpha.app/s/abridge-used-upmc-funding-and-data-to-build-its-early-ai/; the episode from there: https://realalpha.app/v/W4_aRWsHogM/?t=724

## Quote

> “I'm obsessed with this problem with my co-founders and you're going to give us a check and you're also going to **give us all your data**.”
> — [Shiv Rao](https://realalpha.app/builders/shiv-rao/) at [16:13](https://realalpha.app/v/W4_aRWsHogM/?t=973) · Upstarts

## Takeaways

1. Abridge began with transformer-based models before LLMs.
2. Rao framed the work as research to secure funding and data.
3. Annotations helped the company build an MVP and attract researchers.

## About this segment

Shiv Rao describes building with transformer models before LLMs, then securing UPMC funding and de-identified data for annotations. That access helped Abridge build an MVP and recruit technical talent.

From “Abridge’s Shiv Rao: Taking On Microsoft To Cure Doctor Burnout With AI”, Jul 23, 2026. [Watch the full episode](https://realalpha.app/v/W4_aRWsHogM/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[12:08](https://realalpha.app/v/W4_aRWsHogM/?t=728) off in the mid 2010s. And, you know, even in terms of our timing of starting this company was three months after attention is all you need. I think that there's like three vintages of AI native company there are post-transformer paper pre-llm companies post-llm pre-agent companies and then now post-agent companies and the name of the game is to be the latest variant as fast as you possibly can. And it's not just in terms of the product that you're building, but also in terms of the way that you operate as a company. So we were post-Transformer.

[12:40](https://realalpha.app/v/W4_aRWsHogM/?t=760) We started with BERT and BioBERT and Longformer and Pegasus and T5. We were fine-tuning all of these Transformer-based models. And then all those skill sets turned out to be incredibly valuable. We just announced our relationship with NVIDIA, and we're building a foundation model at the conversational level for healthcare. But all those skill sets, we still apply to this day. But when LLMs came out, obviously, you refactor very, very quickly what the product can look like, what the technology can look like, how fast you can go, how fast you can scale. And now in this agent moment, it's just an entirely different game as well. And I think the companies that are going to win are going to be the ones that are super agile, because the ground's going to continue to shift on us we've talked to people on the show like may habib who was you know working

[13:28](https://realalpha.app/v/W4_aRWsHogM/?t=808) with the transformer basically as soon as the paper comes out and then i've also you know talking to people like uh we both know the ceo of zoom eric yuan and eric was telling me recently that he actually beats himself up a little bit that if he's honest he didn't jump on llms and sort of the ai ai potential until that chat chpc moment early on were you betting a lot the company on this technology and and i guess how much was a tech moat going to be you know key to a bridge at least in the early days yeah we placed a really big bet that this was going to work that somehow we were living we're at this intersection of moments in technology and also in health care and on the tech side like deep investments and annotations like when we started of the company, we were finding ways, super creative ways, because early stage startup,

[14:20](https://realalpha.app/v/W4_aRWsHogM/?t=860) its rounds weren't what they look like today in those seed rounds. But we absolutely felt like outsiders, like we were trying to do something that in so many ways was going against the grain. Number one, just in terms of the product, recording conversations, not normal in healthcare with HIPAA and all the other privacy issues, really like going against the grain. This thesis about AI and healthcare taking off, not, you know, had not taken off. And so this idea that the conversation could be computable, that you could build a platform on top, definitely something that investors had a hard time sort of grappling with. But then from a technology standpoint, really believing that over time this this this this would unlock and at the time you know pre-llm and and pre this moment that we're in the the name of the game was fine-tuning these models but it involved

[15:14](https://realalpha.app/v/W4_aRWsHogM/?t=914) lots of really expensive annotations so we were getting really creative finding ways to get non-dilutive capital to employ a bunch of doctors and nurses to go through data sets and create annotations that could allow us to even get to the contours of an MVP. So yeah, it was a very different game, but I think it's helped us. Can you give me an example of what a creative way to find capital would look like? Yeah, it was like going to UPMC, for example, and trying to frame the challenge as research and trying to find a way to get them to foot the bill on a data that they could help, that they could use as well, but that we would obviously find a way to leverage. As a corporate VC, I one day just went to my boss at UPMC, a president over there, and just said, like, I quit. I'm going to start this company. I didn't

[16:02](https://realalpha.app/v/W4_aRWsHogM/?t=962) want it to be a spin out. And I just wanted to make sure that we can own and control that P&L so that we could, you know, control our own destiny. And I also said at the moment, like, hey, like, I'm going to quit. I'm going to start this company. I'm obsessed with this problem with my co-founders and you're going to give us a check and you're also going to give us all your data. And then we're going to annotate as much of that data as possible and we're going to build something. So it was a very YOLO moment, but we got there. And I look back at those emails, I was looking back at them recently and I'm not sure exactly how it all worked out, but it did work out. And they agreed to that? Yeah, they did. They gave the data? They gave us a lot of de-identified data in a responsible way. But without that data, we couldn't be an AI company

[16:51](https://realalpha.app/v/W4_aRWsHogM/?t=1011) off the bat. Without that data, we couldn't have attracted these professors or PhDs to work with us. So it gave us something that allowed us, even if we weren't capitalized as such, it made us operate like we were a deep tech company, even off the bat.