00:01:09 · Arvind Jain
[00:01:09] S0: Well, I mean, I think part of it is I like doing it. I mean, that's the main motivation. Building a startup from scratch, building a great product that, you know, your friends, your families, everybody use, you know, that's such a great feeling, you know, to do. So it's always the motivating factor for me. Love to work. I love to build products. But the motivation for Glean actually came from while I was working at Rubrik, the company that we started in early twenty thirteen. And so it was a problem that we faced at Rubrik. As we grew, our productivity metrics started to go down, and people were complaining about this specific problem that Glean set out to solve, which is everybody was saying that it's very hard for me to work in the company because there's so much information and data spread across so many different places, and I cannot find anything. I cannot find the information I need to do my work. I also can't figure out who are the people who can help me get that work done. So this was a problem that our team was complaining a lot about. And my background, as you were mentioning, I've been a search engineer for most of my engineering life. And so when people said, hey, they can't find things inside the company, I said, why not? Why don't we have a Google like thing inside the company? And it turned out that there was no product in the market for that. And so that's what led to the creation of Glean.
00:02:22 · Guillaume Moubeche
[00:02:22] S1: Super interesting. And I think whenever we think of search now and we connect it to AI, one of the big problem that we see is like models hallucinating. How exactly do you manage to kind of like solve the issue with Glean?
00:02:38 · Arvind Jain
[00:02:38] S0: Yeah. So I think maybe just for quick background. So Glean is a coworker, enterprise coworker. So you can use it inside your company. Typically, the company deploys Glean for all the employees all at once. And when they deploy Glean inside the company, they connect Glean with all of their enterprise systems. And so that way, Glean is now able to understand your business, is able to see all the data and all the knowledge that lives inside your company. It's also able to understand who the different people are, your employees, your customers. It has that context of how your business actually works. And once we have it, now you actually offer our coworker to each one of your employees. And that coworker will look like any other AI coworker that you see today, whether it's Claude or ChatGPT or Gemini. Glean also looks the same. But the main difference is that Glean has, number one, access to all the great AI models that are out there, closed domain models, open source models. So it feels like a superset of ChatGPT or Claude or Gemini all combined into one. But more importantly, because it's connected so deeply and broadly to all of your enterprise systems, it knows how your company works. It has that context. So it can bring the right context, the right raw materials to AI. So when you ask Glean to do work, it's able to use that context and the power of models to do effective work for you. So it's sort of... It it works great. And then Glean is also an AI platform. So the fact that we are connecting with all of enterprise systems and we can build this really rich context and knowledge graph layer, we actually offer that also to other AI tools. So not just our own coworker or our own agent builder, But if you're building agents elsewhere, you can also take advantage of clean... Glean enterprise context. Now... But this context is the key. So we know you were talking about hallucinations. And you ask AI to do things and can make mistakes. It makes... It still makes a lot of mistakes. It's actually shocking how models are still very sort of unpredictable. Even though they're getting better, they can run for a long time, they can do fancy things, they still fundamentally make mistakes. And the way to solve for that is basically by really building good context inside your enterprise.
00:04:52 · Arvind Jain, Guillaume Moubeche
So when you ask AI to do work for you, you have to set the right guardrails for AI. You have to, given any task, first without using AI, generate capabilities. You want to make sure that for any given task that you're assembling the right raw materials. The right context, the information, the data that could solve that task. And so we do that. So we actually assemble the right context for whatever task that you're trying to do. Then we will ask the models to say that, hey, do this work, but use this context. This is your raw material, so make sure to use it to do your work. So we sort of put the right guardrails, the right prompts into the model so that the models stay in the right direction. But once the models produce the work and they do something back, we do validation. We will make sure that every line of response from the AI model, we back check that against the context that we provided. And if we don't see evidence of that information that the model is summarizing for us, then we will redo the work. So that's sort [00:05:52] S1: of like when you one way evidence, you want you want like quotes from documents or this kind of thing? Like, do you do text matching or... Yeah. Okay.
00:05:52 · Guillaume Moubeche, Arvind Jain
[00:05:52] S1: of like when you one way evidence, you want you want like quotes from documents or this kind of thing? Like, do you do text matching or... Yeah. Okay. [00:06:01] S0: Exactly. Yeah. In fact, like every line, like, you know, when Glean gives you an answer, for example, if you're doing a question answering use case, every line of... In that response, we will also put a reference to the authoritative up to date information from within your enterprise. If you cannot find it, then we are not willing to show it to the user. So this is like, you know, basic engineering. Right? Like, you know, you you work with the models, you know they're gonna make a mistake, so you have a layer on top to sort of do that validation. [00:06:28] S1: And also like you're working with a lot of enterprise, which means like large scale contracts, etcetera. We've seen in the last years, you know, like the struggle for some enterprise, you know, to trust the models, like we've seen, you know, like leaks, something like happening between Apple and ChatGPT, etcetera, etcetera. Like how exactly do you balance this for your enterprise clients to make sure that, you know, their data still stay their data, that the confidential information stay confidential? Because the moments you start using, you know, like whether it's Anthropic or OpenAI, eventually the data can still be like used, you know?
00:07:09 · Arvind Jain
[00:07:09] S0: Yeah. It's a very good question. We take a lot of effort to ensure that we are protecting the confidential data and information that an enterprise has. In fact, our working model, first of all, is that every information in the enterprise is confidential to them, not some. Like, you know, it's everything. Right? Like, you know, it's all theirs. So first, like, Glean does work with different model providers, and we have the right arrangements and the contracts with them to ensure that, you know, we are offering zero day data retention capability to our customers. So we ensure that any information that we use from the enterprise and have model work on it, that the model provider cannot actually record that, save it on the desk. We are strictly offering zero day retention. But then that does rely a little bit on the third parties, the model providers, to do the right things. There can also be mistakes or bugs. Engineering systems always can have problems. So you have to go above and beyond. So yes, configure zero day retention. But then when we are doing work with AI in the enterprise, we have also a lot of mechanisms to ensure through runtime security that we are, number one, not taking sensitive information, sending it as is to the model. It's like, you know, we can always do some level of obfuscation, some level of PII remapping to sort of make sure that even if there's a mistake on the model side by the provider, and although they promise us zero day retention, they actually don't do it, we never actually send the sensitive information to them in the first place. Because we kind of obfuscated it into a packet which is safe. And so that's an example of a technique that we will place to further help our enterprise customers feel comfortable with using AI. And that's important because we work with the world's largest enterprises with lots and lots of sensitive information. And then finally, we also offer model choice. So we can actually... For every enterprise, we can actually come up with a list of models that they feel comfortable with and just use them. That's the advantage of being model agnostic and have like 100 plus models in our model hub. And this includes like self hosted models.