theCUBE + NYSE Wired: Mixture of Experts Series | Aabhas Sharma, Hebbia

More from theCUBE + NYSE Wired: Mixture of Experts Series

Aabhas Sharma

CTO

Hebbia

Info

Aabhas Sharma, Hebbia

Aabhas Sharma

CTO Hebbia

search

Aabhas Sharma

>> Palo Alto Studio connecting Silicon Valley and Wall Street. I'm John Furrier, the co -host of the studio here with Dave Vellante, my co -host.

Gemma Allen

>> Welcome back to theCUBE Studio here at the New York Stock Exchange. I'm Gemma Allen with NYSE Wired's Mixture of Experts, where we are connecting Silicon Valley to Wall Street. Join me now for a conversation on how AI is doing a lot more than just wrangling data, it is actually building the infrastructure behind iconic powerhouses like this one. It's Aabhas Sharma, CTO of Hebbia. Welcome, Aabhas.

Aabhas Sharma

>> Thank you for having me.

Gemma Allen

>> So you guys have a very interesting partnership with ICE. I want to get to that, but first, tell me exactly what it is that Hebbia does.

Aabhas Sharma

>> Yeah, so Hebbia is the platform for institutional finance. Think of the world's largest asset managers in the world's largest investment banks. They use Hebbia to run research to do diligence, to build end -to -end workflows. It's really where they come and code their firm -wide way of thinking and their firm -wide way of making decisions into a platform where it can be repeated at scale and really supercharge a lot of their teams. So if you think about the mechanical work that goes into institutional finance, we help them take away the boring mechanical work and let them focus on the final judgment call, the final decision -making, and really increase the deal throughput that they can actually handle.

Gemma Allen

>> So you guys have a relationship with ICE. They are a customer of yours. Maybe use them as a use case. Talk to us about how they are implementing this technology every day.

Aabhas Sharma

>> Yeah, so our customers obviously use ICE for institutional grade market data. Until today, they were switching platforms or switching terminal windows or switching browsers to get access to that data from ICE. and really with a partnership with ICE, that data is now available in the workflows within Hebbia. So they don't have to switch context. They don't have to think about a different tool to open. And really that data compounds for them. So if they're chatting over that data, if they're using it in a workflow within Hebbia, they can just continue to keep building and it just increases this floor, raises this floor for them to keep doing more work, to keep doing more research. And so that institutional grade data is something they were using through different platforms and it's now all under one umbrella so that they don't have to ever leave their workflow. and can deliver results or outcomes from within one platform.

Gemma Allen

>> So it creates, I guess, a simplified, integrated, optimized version of what has been very siloed spaces, right?

Aabhas Sharma

>> Yeah, and it's also highly citable. So maybe I should call that out, which is, at the end of the day, you don't want to make financial decisions without a high level of trust, a high level of auditability. So it always has to be auditable back to the source of truth. So we're not taking the ICE data and changing it. It is the ICE data coming up in the workflows that you're already building, but it's always citable back to the ground truth. So you can always see where that data came from and you trust the actual data because it came from ICE.

Gemma Allen

>> Well, everyone needs a throat to choke, right, when things go wrong, but that's important. So talk to me about companies like this, the space you're in, the competitive dynamic of this industry. What I'm interested in too is the build versus buy dilemma, right? On one side, we hear it's so easy to vibe code solutions. Everyone's a technologist now. We're going to just have a mass influx of technology geniuses across all organizations. And then on the other side, we see companies like this that are clearly meeting a need that internally can't be met. Why would an institution like ICE not build this themselves? Give me the pitch on that.

Aabhas Sharma

>> So if you think about it, our users rely on ICE data, but they rely on Fitch or Moody's or S&P. There's a lot of different data streams. There's a lot of different entitlements. Some firms have access to ICE. Some firms have access to Cap IQ. Some firms have access to FactSet. Those entitlements need to be respected. So that's number one, which is just one platform, building it for themselves and just giving access to that one data source doesn't actually cover the entire sector that we're servicing. So that's part one. Part two is individual productivity will become table stakes or has already become table stakes. If you think about ChatGPT, if you think about Claude, we expect that every computer in the next few years will just come pre-installed with one of those two. And what that does is it makes the individual be far more productive. You can write emails faster. You can go through your emails and summarize them faster. But it doesn't make the entire firm more productive. And that's really where tools like ours come in, which is we're hyper -focused on one vertical, and we're trying to make the entire firm be a lot more productive. So taking all of the things that have lived in analysts' heads or have lived on pieces of paper or live across discrete documents and really institutionalizing it, making it repeatable. And then they don't have to worry about the model layer. We think the model layer is going to keep getting smarter. We think inference is going to keep getting cheaper. But there's always going to be this race on what is the better model today. Some people might like Claude Opus today. Some people might like GPT. Some people might like an open weights model. With our platform, it's our job to figure out what is the best model for the actual task that our consumers are trying to do and to figure out how do we deliver them the ideal outcome. And sometimes that doesn't require the latest and greatest model. So our team is responsible for figuring out which models chained together will actually deliver the actual end result. And how do we move away from individual productivity to making the entire firm be more productive, to make the entire firm be more profitable at the end of the day?

Gemma Allen

>> This is true solution selling. How is it sold? Is it usage? Is it seat based? Is it outcome based? What's the business model behind this?

Aabhas Sharma

>> Yeah, it's a very loaded question. it started off as seat -based, but it's moved more and more towards consumption -based pricing and outcome-based pricing.

Gemma Allen

>> Which is the industry generally, right? Everybody is saying the same thing.

Aabhas Sharma

>> 100%. And that also influences our sales motion, right? So if you remember maybe a year ago, there was a study that came out about how enterprise AI tools are failing at adoption.

Gemma Allen

>> 70 %?

Aabhas Sharma

>> Yeah.

Gemma Allen

>> That's the number?

Aabhas Sharma

>> it's come up in every interview that I've been in. And the reason they were failing, I think, was because they demo really well but at the end of the day you were just giving someone it's the equivalent of if someone gives you the keys to a Ferrari it's not like you're going to get the actual value of a Ferrari out of it unless you're a race car driver right sometimes if you've driven a Prius all of your life and I give you the keys to a Ferrari you're still going to drive it like a Prius it's the enablement to teach you how to drive the Ferrari that will actually make you get the value out of it so even our sales motion has to match that outcome-based deliverable or that outcome-based pricing, which is you have to actually enable these folks. So we have a large team of forward-deployed bankers, forward-deployed investors. These are ex-professionals from the exact firms that we sell into. Some of them were even Hebbia Power users at those firms that then come join Hebbia and really enable these organizations to build those repeatable processes to actually get to those end outcomes in that firm-specific way of doing things. And that makes selling easier. That makes pricing it easier because now you're pricing on outcomes, you're pricing on consumption, you're not pricing on who's using which individual seat, because that's how individual productivity tools should price. But for us, where we're delivering firm -wide transformation, where we're delivering firm -wide deliverables, outcomes is the correct way to price it.

Gemma Allen

>> So who are you competing with in this space? And it's kind of net new-ish, because I'm sure for some of your customers, they probably weren't really doing anything like this up until AI became the mass available option it is today? Or are there other players? Are you competing with other harness players, for example, like Anthropic? Or who is the true competitive force here?

Aabhas Sharma

>> I actually think the answer I'll give you is less that there's a company that's a competitive force here, but it's more that the in-house build is the competitor. Because there are some firms that have massive internal tech teams that will often try to build a solution in-house. And unfortunately, we have to wait to see how that goes. And oftentimes, it ends up failing because that's not their core focus, right? Their core focus has been building out trading systems or building out internal data warehouses and data lakes to make their data be more accessible. But building this repeatable workflow layer that is using the best model, whether it's a foundation model or an open weight model, is not their thing that they've always worked on. So the more frequent competitor ends up being a firm trying to build it in-house and then eventually realizing that they should just let a vertical specific layer, i .e. at Hebbia, actually build it for them or just sell it to them. We don't end up competing with Anthropic or OpenAI. We have very strong partnerships with both. We love working with both. And as the model layer gets better, our product gets better. So it's more of a better together story than a competitive story with both the model layers.

Gemma Allen

>> Because you are a selection vendor in some respects for those models, right? Let's talk about the models though. And let's talk about what percentage of what you do build is proprietary versus built on these models or just sourcing and optimizing the usage of these models across enterprise workflows. Break that down.

Aabhas Sharma

>> Yeah. I'll maybe take a step back and say models have always relied on context windows. Context windows are getting much larger. But if you look at how our customers use our product, they're usually querying over thousands or tens of thousands of documents, which is still too large for the existing context window. Even if the context window gets 10x larger in the next year, when you give it a full VDR or full deal room, it just can't fit in the context window. And RAG by design will pick snippets from these documents to answer your question. And at that point, you're relying on luck to make sure it's picking the right snippet. there's some optimizations you can do to make sure that it picks the right snippets. Whereas on our side, what we do is when all these private documents are ingested, we actually pre-index them, extract all the key terms and build an internal knowledge graph for the organization. So that isn't really using a model layer. That's using our proprietary indexing layer, our proprietary ingestion layer to go extract from each of these documents the fields that actually matter to our ICP, which is why it's very much a vertical specific solution. You unfortunately can't just take the solution and apply it to pharma research today. And that's because the actual content from these documents is critical. Once we've actually indexed all that information, when you start asking questions, we know exactly which documents will contain the pertinent answers. And then we can use the context window. So it's before even the model layer comes in, there's a lot of work that happens to actually make the model layer do the actual purpose that it can do really well. And then when you come to the model layer, we have our own internal orchestration, which is, Claude Opus or GPT-5 isn't needed to answer a lot of these questions, and nor is it economically viable to answer a lot of these questions. So a lot of what we do is invest in open weight models that have been post-trained to answer the exact questions that our ICP has. And oftentimes we'll be routing to these much cheaper models where the cost of inference is just significantly lower. It's only when you need deep reasoning multiple turns multiple thought processes where a foundation model gets pulled in at the end of the day we are building it to be model agnostic it gives us obviously economic leverage to be model agnostic but it also gives our end customer the ability to pick the model that they prefer but a lot of the actual plumbing and orchestration we don't even expose the model layer to the end customer because it's some of it is deterministic and then some of it is going through open weight models.

Gemma Allen

>> So you are friends with everyone. You are the Switzerland of infrastructure of capital markets, which is a great place to be. But let's go back to open weight models for a second. So interesting week with GLM-5.2 and even Fugu, who might be somewhat of a competitive option in some respects to what you do, correct? Talk me through that. what are you seeing from the perspective of open weights? Because we know their cost per token is lower. We know there's advantages. There's also this other narrative that no CISO is going to deploy an openweight model because of the geopolitical risk across their stack. What do you think? And also comment on Fugu. I want to know what your thoughts are on that as a competitive option to what you deliver.

Aabhas Sharma

>> I'll maybe start with a competitive option, which is there's going to be a lot of harness layer companies that will come up. There will be some orchestration layer companies that will come up. A lot of them will be purposely broad because the actual economics that they're optimizing for is token use. Whereas we are very specific, very narrowed in on one vertical. We're not trying to be broad. We're trying to be very opinionated. And that's because the way you process a CIM and read a CIM at one firm versus how you read a CIM at a different firm is very consistent. So instead of trying to boil the ocean, we're trying to just boil the ocean for one vertical. When it comes to open weight models, we've seen really, really good success with post training on open weight models, especially given the fact that we're, again, focused on one vertical. So we're getting in the high 90 percentage success rates close to a foundation model with open weight models now. And again, deploying it on B200s, deploying it on GPUs that we're purchasing through inference providers. It lets us actually get better economics because we're getting sustained usage. It lets us plan for peaks that might come up with a big deal. So SpaceX is about to IPO and suddenly every single one of our customers is working on this IPO. and you know that there's going to be a lot more demand, we can actually prepare for those kinds of spikes. So it just lets us be a lot more in control of our own destiny with these open weight models. And as long as we're deploying on US infrastructure with all the correct safety rails in place, we've not actually had much pushback from CISOs to date.

Gemma Allen

>> Interesting. So all good startups, you guys are scrappy, right? I love that. I love a scrappy story. So I am a new customer, a new bank, a new hedge fund. And I want to use Hebbia, talk about the deployment, the time to value, what does it look like? One thing that I'm interested in that you mentioned is, there's thousands of documents, right? And we know that discoverability has been a long-term challenge for a lot of large financial institutions or lots of different verticals. How do you solve for that? Break down the process to onboard?

Aabhas Sharma

>> Yeah, multi-step answer. So even before the contract is signed, there's a lot of work that goes into security, compliance, and diligencing, how we deploy infrastructure, how we build individual tenants. From day one, we've invested very heavily in what we would call the enterprise ecosystem. So making sure that we work natively or seamlessly with Microsoft SharePoint, with Box, with any of the ways that you might store your documents within your firm, including actually integrating with firm-specific document storage solutions. A lot of these firms that have been around for a while have actually built bespoke document storage solutions in-house.

Gemma Allen

>> And the second part is that they're often not governed, though, correct? That they're often not governed, that often are just massive file shares in some cases, right? So there's a lot of homework in that, I'm sure.

Aabhas Sharma

>> Oh, 100%. So that's why there's a lot of scoping that gets done up front, because part of it is actually figuring out how much work it will be for us to integrate into your on-prem data storage solution or into your bespoke document storage solution. they also oftentimes will have their own bespoke way of doing permissions. there's the Oktas of the world, which are off-the-shelf solution systems, but a lot of the large banks have built entitlements and permissioning in-house that tracks, which team has access to what documents, which team has access to what data sources. So part of it is scoping out all of that and understanding how that will hook into our system so that their entitlements flow over, so that their user permissions flow over, so that their documents flow over, and understanding the size of the document corpora that they have. And then convincing the CISO that, hey, our individual tenant that we're going to build for you, where you can encrypt it with your own key, will make it so that none of our employees can see the data you're bringing. Because that is very important to our customers. They have material non-public information, and they want to make sure that no one else can see it. So we've invested very heavily in building out those compensating controls to make the CISOs, the CIOs comfortable with our technology. We also have really strong agreements in place to make sure the model layers never train on our customers' data. And then obviously with the open weight models, we're again not training on our customers' data. We have really strict policies in place around data retention and making sure that it is something that the firm will be okay with. So all of that is the pre -work. Once the contract is signed, there's usually a few weeks where we will then be integrating into their document solutions, into their SharePoint, into their Box folders, into their permissioning systems, syncing over all those permissions. And then really it's our forward deployed expertise, right? Which is these ex -bankers, these ex -investment professionals who work for Hebbia, who are experts in our primitives and our Lego blocks, if you really want to simplify it, that sit alongside your finance professionals, alongside your investment bankers, understand the process, whiteboard the process that they have and encode it into our system. And so then really at the end of the day, we're not giving you the keys to a Ferrari. We're teaching you everything about driving that Ferrari. We're teaching you how to get the most value out of the Ferrari before you ever get into the driver's seat so that when you get into the driver's seat, you can actually get the most value out of it.

Gemma Allen

>> I love the pragmatism of that because often, I talk to a lot of startups and I often hear, how quickly can you get value from this? Oh, immediately, today? And the reality is we know in tech, anything that gives you value that quickly probably should arrive with some level of suspicion, right? There is often effort involved in streamlining something that's been highly complicated and highly fragmented for such a long time. So great answer. Last question, the company. Talk to me about the stage you guys are at, financials, funding, and what's ahead? What's on the roadmap? Where do you go from here?

Aabhas Sharma

>> Yeah, so we raised our Series B almost two years ago. We're really fortunate to have Andreessen, Index Ventures, Google Ventures on our cap table. We're really well capitalized. Company has been growing really fast. We were about 50 or 60 folks a year ago. We'll probably end this year around 300 folks. So really exciting time. And there's really been this inflection point in our usage as well. Traditionally, we used to see contracts with 100 or 200 seats. But one of the largest bulge bracket banks in the world just signed a contract with us to roll out 6 ,500 seats. So the contract sizes are changing to really be these wall -to -wall deployments where every single person in the firm is starting to get access. So we're staffing up to really meet that demand and make sure that they get the most value as quickly as possible. And yeah, super exciting time to be at the company and to continue growing.

Gemma Allen

>> Love it. Staying scrappy, hiring and blazing a trail. Aabhas Sharma, thank you so much for joining us on theCUBE.

Aabhas Sharma

>> Thank you for having me.

Gemma Allen

>> I'm Gemma. I'm here at theCUBE studio at the New York Stock Exchange.This is NYSE Wired's Mixture of Experts. Thanks for watching.

Aabhas Sharma, Hebbia

search

Aabhas Sharma

>> Palo Alto Studio connecting Silicon Valley and Wall Street. I'm John Furrier, the co -host of the studio here with Dave Vellante, my co -host.

Gemma Allen

>> Welcome back to theCUBE Studio here at the New York Stock Exchange. I'm Gemma Allen with NYSE Wired's Mixture of Experts, where we are connecting Silicon Valley to Wall Street. Join me now for a conversation on how AI is doing a lot more than just wrangling data, it is actually building the infrastructure behind iconic powerhouses like this one. It's Aabhas Sharma, CTO of Hebbia. Welcome, Aabhas.

Aabhas Sharma

>> Thank you for having me.

Gemma Allen

>> So you guys have a very interesting partnership with ICE. I want to get to that, but first, tell me exactly what it is that Hebbia does.

Aabhas Sharma

>> Yeah, so Hebbia is the platform for institutional finance. Think of the world's largest asset managers in the world's largest investment banks. They use Hebbia to run research to do diligence, to build end -to -end workflows. It's really where they come and code their firm -wide way of thinking and their firm -wide way of making decisions into a platform where it can be repeated at scale and really supercharge a lot of their teams. So if you think about the mechanical work that goes into institutional finance, we help them take away the boring mechanical work and let them focus on the final judgment call, the final decision -making, and really increase the deal throughput that they can actually handle.

Gemma Allen

>> So you guys have a relationship with ICE. They are a customer of yours. Maybe use them as a use case. Talk to us about how they are implementing this technology every day.

Aabhas Sharma

>> Yeah, so our customers obviously use ICE for institutional grade market data. Until today, they were switching platforms or switching terminal windows or switching browsers to get access to that data from ICE. and really with a partnership with ICE, that data is now available in the workflows within Hebbia. So they don't have to switch context. They don't have to think about a different tool to open. And really that data compounds for them. So if they're chatting over that data, if they're using it in a workflow within Hebbia, they can just continue to keep building and it just increases this floor, raises this floor for them to keep doing more work, to keep doing more research. And so that institutional grade data is something they were using through different platforms and it's now all under one umbrella so that they don't have to ever leave their workflow. and can deliver results or outcomes from within one platform.

Gemma Allen

>> So it creates, I guess, a simplified, integrated, optimized version of what has been very siloed spaces, right?

Aabhas Sharma

>> Yeah, and it's also highly citable. So maybe I should call that out, which is, at the end of the day, you don't want to make financial decisions without a high level of trust, a high level of auditability. So it always has to be auditable back to the source of truth. So we're not taking the ICE data and changing it. It is the ICE data coming up in the workflows that you're already building, but it's always citable back to the ground truth. So you can always see where that data came from and you trust the actual data because it came from ICE.

Gemma Allen

>> Well, everyone needs a throat to choke, right, when things go wrong, but that's important. So talk to me about companies like this, the space you're in, the competitive dynamic of this industry. What I'm interested in too is the build versus buy dilemma, right? On one side, we hear it's so easy to vibe code solutions. Everyone's a technologist now. We're going to just have a mass influx of technology geniuses across all organizations. And then on the other side, we see companies like this that are clearly meeting a need that internally can't be met. Why would an institution like ICE not build this themselves? Give me the pitch on that.

Aabhas Sharma

>> So if you think about it, our users rely on ICE data, but they rely on Fitch or Moody's or S&P. There's a lot of different data streams. There's a lot of different entitlements. Some firms have access to ICE. Some firms have access to Cap IQ. Some firms have access to FactSet. Those entitlements need to be respected. So that's number one, which is just one platform, building it for themselves and just giving access to that one data source doesn't actually cover the entire sector that we're servicing. So that's part one. Part two is individual productivity will become table stakes or has already become table stakes. If you think about ChatGPT, if you think about Claude, we expect that every computer in the next few years will just come pre-installed with one of those two. And what that does is it makes the individual be far more productive. You can write emails faster. You can go through your emails and summarize them faster. But it doesn't make the entire firm more productive. And that's really where tools like ours come in, which is we're hyper -focused on one vertical, and we're trying to make the entire firm be a lot more productive. So taking all of the things that have lived in analysts' heads or have lived on pieces of paper or live across discrete documents and really institutionalizing it, making it repeatable. And then they don't have to worry about the model layer. We think the model layer is going to keep getting smarter. We think inference is going to keep getting cheaper. But there's always going to be this race on what is the better model today. Some people might like Claude Opus today. Some people might like GPT. Some people might like an open weights model. With our platform, it's our job to figure out what is the best model for the actual task that our consumers are trying to do and to figure out how do we deliver them the ideal outcome. And sometimes that doesn't require the latest and greatest model. So our team is responsible for figuring out which models chained together will actually deliver the actual end result. And how do we move away from individual productivity to making the entire firm be more productive, to make the entire firm be more profitable at the end of the day?

Gemma Allen

>> This is true solution selling. How is it sold? Is it usage? Is it seat based? Is it outcome based? What's the business model behind this?

Aabhas Sharma

>> Yeah, it's a very loaded question. it started off as seat -based, but it's moved more and more towards consumption -based pricing and outcome-based pricing.

Gemma Allen

>> Which is the industry generally, right? Everybody is saying the same thing.

Aabhas Sharma

>> 100%. And that also influences our sales motion, right? So if you remember maybe a year ago, there was a study that came out about how enterprise AI tools are failing at adoption.

Gemma Allen

>> 70 %?

Aabhas Sharma

>> Yeah.

Gemma Allen

>> That's the number?

Aabhas Sharma

>> it's come up in every interview that I've been in. And the reason they were failing, I think, was because they demo really well but at the end of the day you were just giving someone it's the equivalent of if someone gives you the keys to a Ferrari it's not like you're going to get the actual value of a Ferrari out of it unless you're a race car driver right sometimes if you've driven a Prius all of your life and I give you the keys to a Ferrari you're still going to drive it like a Prius it's the enablement to teach you how to drive the Ferrari that will actually make you get the value out of it so even our sales motion has to match that outcome-based deliverable or that outcome-based pricing, which is you have to actually enable these folks. So we have a large team of forward-deployed bankers, forward-deployed investors. These are ex-professionals from the exact firms that we sell into. Some of them were even Hebbia Power users at those firms that then come join Hebbia and really enable these organizations to build those repeatable processes to actually get to those end outcomes in that firm-specific way of doing things. And that makes selling easier. That makes pricing it easier because now you're pricing on outcomes, you're pricing on consumption, you're not pricing on who's using which individual seat, because that's how individual productivity tools should price. But for us, where we're delivering firm -wide transformation, where we're delivering firm -wide deliverables, outcomes is the correct way to price it.

Gemma Allen

>> So who are you competing with in this space? And it's kind of net new-ish, because I'm sure for some of your customers, they probably weren't really doing anything like this up until AI became the mass available option it is today? Or are there other players? Are you competing with other harness players, for example, like Anthropic? Or who is the true competitive force here?

Aabhas Sharma

>> I actually think the answer I'll give you is less that there's a company that's a competitive force here, but it's more that the in-house build is the competitor. Because there are some firms that have massive internal tech teams that will often try to build a solution in-house. And unfortunately, we have to wait to see how that goes. And oftentimes, it ends up failing because that's not their core focus, right? Their core focus has been building out trading systems or building out internal data warehouses and data lakes to make their data be more accessible. But building this repeatable workflow layer that is using the best model, whether it's a foundation model or an open weight model, is not their thing that they've always worked on. So the more frequent competitor ends up being a firm trying to build it in-house and then eventually realizing that they should just let a vertical specific layer, i .e. at Hebbia, actually build it for them or just sell it to them. We don't end up competing with Anthropic or OpenAI. We have very strong partnerships with both. We love working with both. And as the model layer gets better, our product gets better. So it's more of a better together story than a competitive story with both the model layers.

Gemma Allen

>> Because you are a selection vendor in some respects for those models, right? Let's talk about the models though. And let's talk about what percentage of what you do build is proprietary versus built on these models or just sourcing and optimizing the usage of these models across enterprise workflows. Break that down.

Aabhas Sharma

>> Yeah. I'll maybe take a step back and say models have always relied on context windows. Context windows are getting much larger. But if you look at how our customers use our product, they're usually querying over thousands or tens of thousands of documents, which is still too large for the existing context window. Even if the context window gets 10x larger in the next year, when you give it a full VDR or full deal room, it just can't fit in the context window. And RAG by design will pick snippets from these documents to answer your question. And at that point, you're relying on luck to make sure it's picking the right snippet. there's some optimizations you can do to make sure that it picks the right snippets. Whereas on our side, what we do is when all these private documents are ingested, we actually pre-index them, extract all the key terms and build an internal knowledge graph for the organization. So that isn't really using a model layer. That's using our proprietary indexing layer, our proprietary ingestion layer to go extract from each of these documents the fields that actually matter to our ICP, which is why it's very much a vertical specific solution. You unfortunately can't just take the solution and apply it to pharma research today. And that's because the actual content from these documents is critical. Once we've actually indexed all that information, when you start asking questions, we know exactly which documents will contain the pertinent answers. And then we can use the context window. So it's before even the model layer comes in, there's a lot of work that happens to actually make the model layer do the actual purpose that it can do really well. And then when you come to the model layer, we have our own internal orchestration, which is, Claude Opus or GPT-5 isn't needed to answer a lot of these questions, and nor is it economically viable to answer a lot of these questions. So a lot of what we do is invest in open weight models that have been post-trained to answer the exact questions that our ICP has. And oftentimes we'll be routing to these much cheaper models where the cost of inference is just significantly lower. It's only when you need deep reasoning multiple turns multiple thought processes where a foundation model gets pulled in at the end of the day we are building it to be model agnostic it gives us obviously economic leverage to be model agnostic but it also gives our end customer the ability to pick the model that they prefer but a lot of the actual plumbing and orchestration we don't even expose the model layer to the end customer because it's some of it is deterministic and then some of it is going through open weight models.

Gemma Allen

>> So you are friends with everyone. You are the Switzerland of infrastructure of capital markets, which is a great place to be. But let's go back to open weight models for a second. So interesting week with GLM-5.2 and even Fugu, who might be somewhat of a competitive option in some respects to what you do, correct? Talk me through that. what are you seeing from the perspective of open weights? Because we know their cost per token is lower. We know there's advantages. There's also this other narrative that no CISO is going to deploy an openweight model because of the geopolitical risk across their stack. What do you think? And also comment on Fugu. I want to know what your thoughts are on that as a competitive option to what you deliver.

Aabhas Sharma

>> I'll maybe start with a competitive option, which is there's going to be a lot of harness layer companies that will come up. There will be some orchestration layer companies that will come up. A lot of them will be purposely broad because the actual economics that they're optimizing for is token use. Whereas we are very specific, very narrowed in on one vertical. We're not trying to be broad. We're trying to be very opinionated. And that's because the way you process a CIM and read a CIM at one firm versus how you read a CIM at a different firm is very consistent. So instead of trying to boil the ocean, we're trying to just boil the ocean for one vertical. When it comes to open weight models, we've seen really, really good success with post training on open weight models, especially given the fact that we're, again, focused on one vertical. So we're getting in the high 90 percentage success rates close to a foundation model with open weight models now. And again, deploying it on B200s, deploying it on GPUs that we're purchasing through inference providers. It lets us actually get better economics because we're getting sustained usage. It lets us plan for peaks that might come up with a big deal. So SpaceX is about to IPO and suddenly every single one of our customers is working on this IPO. and you know that there's going to be a lot more demand, we can actually prepare for those kinds of spikes. So it just lets us be a lot more in control of our own destiny with these open weight models. And as long as we're deploying on US infrastructure with all the correct safety rails in place, we've not actually had much pushback from CISOs to date.

Gemma Allen

>> Interesting. So all good startups, you guys are scrappy, right? I love that. I love a scrappy story. So I am a new customer, a new bank, a new hedge fund. And I want to use Hebbia, talk about the deployment, the time to value, what does it look like? One thing that I'm interested in that you mentioned is, there's thousands of documents, right? And we know that discoverability has been a long-term challenge for a lot of large financial institutions or lots of different verticals. How do you solve for that? Break down the process to onboard?

Aabhas Sharma

>> Yeah, multi-step answer. So even before the contract is signed, there's a lot of work that goes into security, compliance, and diligencing, how we deploy infrastructure, how we build individual tenants. From day one, we've invested very heavily in what we would call the enterprise ecosystem. So making sure that we work natively or seamlessly with Microsoft SharePoint, with Box, with any of the ways that you might store your documents within your firm, including actually integrating with firm-specific document storage solutions. A lot of these firms that have been around for a while have actually built bespoke document storage solutions in-house.

Gemma Allen

>> And the second part is that they're often not governed, though, correct? That they're often not governed, that often are just massive file shares in some cases, right? So there's a lot of homework in that, I'm sure.

Aabhas Sharma

>> Oh, 100%. So that's why there's a lot of scoping that gets done up front, because part of it is actually figuring out how much work it will be for us to integrate into your on-prem data storage solution or into your bespoke document storage solution. they also oftentimes will have their own bespoke way of doing permissions. there's the Oktas of the world, which are off-the-shelf solution systems, but a lot of the large banks have built entitlements and permissioning in-house that tracks, which team has access to what documents, which team has access to what data sources. So part of it is scoping out all of that and understanding how that will hook into our system so that their entitlements flow over, so that their user permissions flow over, so that their documents flow over, and understanding the size of the document corpora that they have. And then convincing the CISO that, hey, our individual tenant that we're going to build for you, where you can encrypt it with your own key, will make it so that none of our employees can see the data you're bringing. Because that is very important to our customers. They have material non-public information, and they want to make sure that no one else can see it. So we've invested very heavily in building out those compensating controls to make the CISOs, the CIOs comfortable with our technology. We also have really strong agreements in place to make sure the model layers never train on our customers' data. And then obviously with the open weight models, we're again not training on our customers' data. We have really strict policies in place around data retention and making sure that it is something that the firm will be okay with. So all of that is the pre -work. Once the contract is signed, there's usually a few weeks where we will then be integrating into their document solutions, into their SharePoint, into their Box folders, into their permissioning systems, syncing over all those permissions. And then really it's our forward deployed expertise, right? Which is these ex -bankers, these ex -investment professionals who work for Hebbia, who are experts in our primitives and our Lego blocks, if you really want to simplify it, that sit alongside your finance professionals, alongside your investment bankers, understand the process, whiteboard the process that they have and encode it into our system. And so then really at the end of the day, we're not giving you the keys to a Ferrari. We're teaching you everything about driving that Ferrari. We're teaching you how to get the most value out of the Ferrari before you ever get into the driver's seat so that when you get into the driver's seat, you can actually get the most value out of it.

Gemma Allen

>> I love the pragmatism of that because often, I talk to a lot of startups and I often hear, how quickly can you get value from this? Oh, immediately, today? And the reality is we know in tech, anything that gives you value that quickly probably should arrive with some level of suspicion, right? There is often effort involved in streamlining something that's been highly complicated and highly fragmented for such a long time. So great answer. Last question, the company. Talk to me about the stage you guys are at, financials, funding, and what's ahead? What's on the roadmap? Where do you go from here?

Aabhas Sharma

>> Yeah, so we raised our Series B almost two years ago. We're really fortunate to have Andreessen, Index Ventures, Google Ventures on our cap table. We're really well capitalized. Company has been growing really fast. We were about 50 or 60 folks a year ago. We'll probably end this year around 300 folks. So really exciting time. And there's really been this inflection point in our usage as well. Traditionally, we used to see contracts with 100 or 200 seats. But one of the largest bulge bracket banks in the world just signed a contract with us to roll out 6 ,500 seats. So the contract sizes are changing to really be these wall -to -wall deployments where every single person in the firm is starting to get access. So we're staffing up to really meet that demand and make sure that they get the most value as quickly as possible. And yeah, super exciting time to be at the company and to continue growing.

Gemma Allen

>> Love it. Staying scrappy, hiring and blazing a trail. Aabhas Sharma, thank you so much for joining us on theCUBE.

Aabhas Sharma

>> Thank you for having me.

Gemma Allen

>> I'm Gemma. I'm here at theCUBE studio at the New York Stock Exchange.This is NYSE Wired's Mixture of Experts. Thanks for watching.