Generative Now | AI Builders on Creating the Future
00:07:28 · Mike Mignano, Gabe Stengel
[00:07:28] S0: Got it. So it almost seems like the thing that really becomes accessible with a tool like Rogo, at least in the early days, was sort of almost like the data science function and like the real deep analytical work. Not necessarily the stuff that, you know, as you said, you can do a lot of this stuff in Chatuchibi Tea now with through like a deep research query. A deep research, I'm sure, isn't doing a lot of the data science work, but it's probably doing some of like the basic just like pulling of of of existing data and records to keep somebody like just generally more high level informed. [00:07:57] S1: I mean, look, Brogo is in many ways like Deep Research except plugged into financial data sources and financial tools. And so instead of only having access to the web and a Python code interpreter environment, our Deep Research has access to the bank's internal data, their CRM, their SharePoint, their precedent transactions analyses. And then it also has access to tools like Excel, like filings, like FactSet, Cap IQ, PitchBook, all the financial data sources that they use. And so it's basically just deep research that's been trained to be a Wall Street analyst as opposed to just anyone's research analyst. Yeah.
00:07:57 · Gabe Stengel, Mike Mignano
[00:07:57] S1: I mean, look, Brogo is in many ways like Deep Research except plugged into financial data sources and financial tools. And so instead of only having access to the web and a Python code interpreter environment, our Deep Research has access to the bank's internal data, their CRM, their SharePoint, their precedent transactions analyses. And then it also has access to tools like Excel, like filings, like FactSet, Cap IQ, PitchBook, all the financial data sources that they use. And so it's basically just deep research that's been trained to be a Wall Street analyst as opposed to just anyone's research analyst. Yeah. [00:08:31] S0: Yeah. So, all right. Let's talk about that a little bit. So you mentioned things like PitchBook. These are, you know, these are data sources that are out there on the web. You have to pay for them, but they're out there. And then you mentioned sort of like the company's data. Is that really like sort of the secret sauce there in that it takes your internal data that's proprietary, and then it links it up with all this stuff that's on the web but maybe hard to get and something like ChatGPT can't just go and grab?
00:08:53 · Gabe Stengel
[00:08:53] S1: I would say there's three things that make Rogo special in our IP and make it a valuable product. One, yes, is the content, the data that you have access to. Right? We work with providers like S and P Global, Cap IQ to make sure that we have, you know, if they... If we have mutual customers, they can query the Cap IQ data through Roko. That is a business data licensing problem. The second is the actual tools that the model has access to. Right? Even if you have this ingenious LLM, if it doesn't have access to a tool like Excel, can it really, you know, build a full three statement model? It can in Python, but that's not auditable and it's not really built for that. Can it make a PowerPoint output? Maybe, but probably not the one that's up to your your output specification. The final thing that's different is our reasoning and the actual model quality. And the same way that, you know, if you look at Claude four or Gemini 2.5 Pro or, you know, GPT four or three, they're ingenious super smart models, but they don't know how to use the tools, the data sets, and do the workflows that bankers and investors do. And if you had, you know, Albert Einstein start at Lightspeed on Monday, just because he's smarter than everyone at light speed, which may or may not be true, doesn't mean he'll actually be a valuable... Maybe. Maybe. I don't think he would be valuable on day one. He needs to learn how you guys think about a company, how you think about investments, how you use tools, what your formatting is, right? And that's actually learning. And so we teach the model how to do that by using a technique called post training or reinforcement fine tuning to learn how to use the tools that finance professionals use.
00:10:28 · Mike Mignano, Gabe Stengel
[00:10:28] S0: So these are your own models. You're not using, you know, GPT? [00:10:33] S1: Well, we build on top of those. So, it's like a lot of these platforms now offer fine tuning on top of their models. So, you can fine tune 4.0. You can fine tune 2.5 Pro. You can also distill these models into open source and then run reinforcement learning. So we're not doing any pre training. But we are then, you know, doing post training on these models. [00:10:51] S0: Yeah. And you mentioned that you're you're sort of you're fine tuning them to to work with the tools that, you know, people in this industry are working with. Talk a little bit about that. Like, do you mean actually sort of agentically using these tools?
00:10:51 · Mike Mignano, Gabe Stengel
[00:10:51] S0: Yeah. And you mentioned that you're you're sort of you're fine tuning them to to work with the tools that, you know, people in this industry are working with. Talk a little bit about that. Like, do you mean actually sort of agentically using these tools? [00:11:02] S1: Oh, Yeah, yeah, If you were an analyst at Goldman and you started on Monday and someone said, hey, find the precedent transactions for the Microsoft Activision deal, you would probably go to EDGAR's, which is the SEC's, you know, list of filings and start looking. Microsoft Activision. Look for the DEF 14 a. Okay. I found it. Now I'm gonna control f through the DEF 14 a for the precedent transactions. Okay. Great. There they are. That's not a tool that an agent has access to. Think we more and more see agents that have browser capabilities and so could That's do that a lot less efficient than just giving an agent access to a filing search tool where it can put in the ticker of a company, filter by a filing type and a date range, and then control F through those documents. And so we both build the tools and then we teach the model how to use it. And so for each different type of tool, we actually can have specialized models that query those documents more effectively. Right? They do the right keywords. Or a specialized tool for, you know, formatting Excels. And so we have to teach the models how to use these in a much lower latency, much higher accuracy sort of way.
00:12:06 · Mike Mignano
[00:12:06] S0: Yeah. I'm sure you're asked this all the time. You know, common question for, I feel like, any AI startup. You know, how do you think about the role of Vrogo in using all these tools and accessing all these data sources that we talked about versus the role of an OpenAI, an Anthropic, you know, an x x AI that, you know, probably want access to many of the same data sources and, you know, maybe over time have the capability to to do, you know, perform many of the same agentic tasks.