1Mby1M Virtual Accelerator AI Investor Forum: With Yash Hemaraj, General Partner at BGV (Part 5)

Sramana Mitra: So I think the subtlety here is that if the decision makers are not adjacent decision makers, then the whole sales cycle has to be repeated. You may get a referral that, “Oh, we’re using this product and this product is doing well for us.” Right now, from a technology point of view, the same technology can be applied to a lot of different things. Your agent technology, for instance, can be applied to various use cases—like marketing, human resources, or engineering—but these aren’t adjacent use cases. They’re far apart within an organization, and in a larger enterprise, that adds complexity.

So I think in general, focusing on all of the boring aspects of one use case and really coming up with a full solution—what Jeffrey Moore refers to as the ‘whole product’—rather than just a point product.

If you have that, and if you train your sales force, marketing, everything to sell that one thing on a repeatable basis, that kind of repeatability is what drives velocity. And without repeatability, if you have to reinvent solutions in every sales cycle, that is not how you can get to velocity.

Yash Hemaraj: I wanted to touch on small, domain-specific models that we discussed earlier. Our thesis is that in vertical workflows, many of these smaller domain models are being developed organically by solution providers focused on vertical industries. For example, in the medical field, experts in regulatory information management are building these models from within.

There’s a balance to strike between leveraging large language models (LLMs) with contextual retrieval-augmented generation (RAG) and building entirely domain-specific models from scratch. This balance is crucial, but it’s still unclear which strategy will dominate, as LLM capabilities are advancing rapidly with increasing context windows.

If a startup aims to build a domain-specific model for healthcare and needs, say, $20 million for a first version, that might be a flawed approach. Instead, if you’re already embedded in the workflow and have access to the necessary data, you can better assess where a small model might serve as a force multiplier and where the rest can be augmented with LLMs. It’s ultimately about system design.

Sramana Mitra: System design is exactly right, because if you can take a large language model that is already trained, and then you can basically deeper train it in a domain and somehow constrain it. And this is where I think there are lots of open questions on how much constraining is possible. And how do you do that? Those are still open questions.

We are intelligent human beings, but we’re not trained to do brain surgery. We train to do venture capital and tech startups, but we’re not trained to do brain surgery. So if you take an intelligent brain and want that brain to do brain surgery, that has to be trained in brain surgery. That’s the, you know, analogy in human versus AI or human applied to AI. But it’s not quite clear exactly how the model is going to happen system design point of view, from a system design point of view.

Yash Hemaraj: Indeed.

Sramana Mitra: Is there any other company that you want to discuss before we wrap up?

Yash Hemaraj: We invest in two  areas. On one side, we focus on enterprise function – software development, customer engagement, revenue operations – being replaced by a co-pilot or a fully autonomous agent. We have individual investments in each of those areas.

On the other side, we have these vertical applications where life is being disrupted by a lot of macroeconomic events, such as supply chain disruptions or climate change. Because of them, you have the opportunity to completely reimagine your workflows. On the agentic software side, I mentioned companies on the vertical applications like AI Dash.

In healthcare, we invest in a company called Essenvia, which automates the entire process of bringing a medical device to market, from creation to regulatory management. For FDA approval, specific forms and answers are required, and the same process must be repeated in Europe and other regions. Essenvia automates each step, and they’ve now expanded to cover post-market filings, including monitoring recalls and submitting clarifications to the FDA. They started with one use case and have expanded into multiple areas within enterprises.

This approach of breaking down industry processes and applying process intelligence allows companies to reimagine workflows entirely. We’re also exploring companies in legal, construction technology, and other sectors. We’re just at the beginning of a wonderful innovation cycle, and it’s exciting to see what’s coming.

This segment is part 5 in the series : 1Mby1M Virtual Accelerator AI Investor Forum: With Yash Hemaraj, General Partner at BGV
1 2 3 4 5 6