Real Alpha
Try Pro free

Upstarts · 3 min · Jul 23, 2026

Abridge routes work between in-house and frontier models

Shiv Rao · CEO, Abridge

Shiv Rao says 40–60% of weekly model output is driven by in-house work, while frontier models remain in the mix.

In this episode

Full analysis
  1. 31:32Abridge challenged Microsoft with a conversation-first vision4 min
  2. 27:00Clinical notes connect care delivery to healthcare payment3 min
  3. 43:21A difficult IVF journey shaped Abridge’s mission3 min
  4. 12:04Abridge used UPMC funding and data to build its early AI5 min
  5. 23:20Abridge started with providers to build a three-sided network3 min
  6. 40:25Abridge routes work between in-house and frontier models3 min
  7. 17:00Abridge moved from a consumer app to enterprise for scale3 min
  8. 20:07IBM Watson lost trust by promising more than it could deliver2 min
“40 to 60 percent of our model outputs are probably driven by in-house work”
Shiv Raoat 41:01 · Upstarts
41:01

Takeaways

  1. In-house work drives 40–60% of weekly outputs.
  2. Frontier models are reserved for selected problems.
  3. Routing and evaluation can shape product quality and P&L.

About this segment

Shiv Rao says 40–60% of weekly model output is driven by in-house work, while frontier models remain in the mix. Routing tasks among models is his path to better products and stronger unit economics.

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.

  • Alex Konrad, Host · Founder & editor, Upstarts Media

    Ran Forbes' Midas List for ten years and co-created its Cloud 100 list; founded Upstarts in 2025.

From “Abridge’s Shiv Rao: Taking On Microsoft To Cure Doctor Burnout With AI”, Jul 23, 2026. Watch the full episode

Transcript

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

  1. 40:29

    We've always been fine-tuning, you know, mid-training now, we're post-training at scale because we have so many users, all those edits and adjustments, that's a proprietary data set that allows us to build a better product, ship a better thing. What the NVIDIA partnership, I think, represents is like, hey, we've got this incredible proprietary data set, what else can we do with it? And it means we can deliver on this promise of 10 agents working for every doctor 24-7.

  2. 40:58

    means we can make it economically viable i mentioned this before in another podcast like 40 on any given week 40 to 60 percent of our model outputs are probably driven by in-house work we still use the frontier models but we try to be very careful about which problems deserve those frontier models and so it's orchestration model routing evals are really the operating system for every single ai company out there and if you do a really good job of that you're going to deliver a better product but you're also gonna like be able to compete with your pnl as well so as these companies like open ai and anthropic go public what's going to happen on the token side on the cost side are companies going to be able to afford are they going to be able to pitch like the big story around all the products that they can deliver are they going to be able to concentrate as much

  3. 41:48

    value in their core product this is where like the the companies that can reach down lower into the stack the ones that can really own and control their destiny are going to be able to separate do you anticipate you'll be competing more with anthropic and open ai over time obviously they've invested in life sciences i think we're going to be partnering with them more and more and i think the moment that we feel like we're going against the grain of of what is absolutely incredible about this moment with how fast things are changing and how profound this platform shift moment is like that if that ever happens like we're screwed like whatever they do needs to feel like a tailwind but it does now like when they do new things Oh, awesome. We evaluate. We see where we should point those models. And then we're delivering a better product experience on some level. But we're also getting all of the edits and adjustments. We're getting the feedback loops. We can decide for which tasks we should end up like distilling or fine tuning or, you know, post training. We can be very careful and deliberate about it.

  4. 42:50

    So we'll always be riding, I think, with them. And we'll also be always investing in, you know, in our own kind of stack as well. And I think that probably it's going to be that sort of orchestration or portfolio approach to models that's going to win the day. Lastly, I know you recently raised a bunch of money. You've gotten more into revenue cycle management, these other areas of the stack. For the average person, for a patient who you want to maybe root for a bridge to succeed, how does this stay helping better outcomes, helping people versus maybe helping us get billed?