The Context Platform for AI Agents: DataHub

DataHub is a context management platform that unifies metadata, lineage, quality, governance, and business knowledge into a trusted context graph for AI agents, data teams, and enterprise applications. With DataHub Cloud and its open source foundation, organizations can keep context accurate, current, and actionable so agents and humans can find, understand, govern, and use data with confidence.

In this episode we explore how DataHub evolved from a LinkedIn data catalog into a vendor-neutral context platform for AI agents. The core idea is that AI needs more than raw data access — it needs complete, trusted, shared context to answer questions accurately. We speak with Shirishanka Das, Co-Founder and CTO of DataHub, and discuss why enterprise AI often fails: models can sound confident while using the wrong definitions or stale data. DataHub’s answer is a graph-based context layer that connects technical, operational, semantic, and business context across structured and unstructured sources, then serves that context to AI agents before they query the data layer.

Follow SourceForge:

SourceForge.net - https://sourceforge.net
SourceForge LinkedIn
SourceForge X (Twitter)
SourceForge Facebook

Interested in appearing on the SourceForge Podcast? Contact us here.

The SourceForge Podcast is the world's largest B2B software podcast.

Transcript

Copy Transcript Copied

SPEAKER_00 0:05

Hello everyone and welcome back to the SourceForge podcast. I'm your host, Bo Hamilton. Now, I want to paint the scene for you for today's episode. Uh, there is a problem just about every company has run into in the last couple of years, especially in the in the AI space here. You ask one of these new AI analytics tools a plain English-based question, like what was our revenue in the Northeast last quarter? And it hands you an answer that's that's fast and it's confident, um, and it's beautifully formatted, but uh, and often often is the case is it's wrong. You know, it grabbed the wrong table. It uh it used maybe marketing's definition of what an active user is instead of the finance definition. And because the number looks right, it it sails right into the slide deck and then it gets factored into a real decision, right? Um, and it turns out this is the sort of defining failure of enterprise AI right now. The models are brilliant, the data underneath them is um is a mess. And it's it's scattered across a lot of different uh systems with definitions that don't agree with one another. And a stronger model doesn't really fix it, it just sort of hallucinates a more convincing answer. Um, so the problem isn't so much intelligence, it's context. And the AI doesn't know what your business actually means by the words it's using. Um, and solving that has become the the sort of the buzzword uh in enterprise data this year. It's that context layer we hear so much about. Um, and this brings us to our guest of the show, Shrashhanka Doss, co-founder and CTO at Data Hub. And Data Hub sits right at the center of this context layer shift. And what's fascinating is they're building it before it had a name. It started as an open source project. Shreshanka created over at LinkedIn to handle GDPR compliance. Then it grew into a community of 15,000 plus contributors and is now a Bessemer uh backed company. Um today they they power data infrastructure at companies like Apple and Pinterest, and they even just launched a context platform purpose-built for AI agents. So they're doing a lot of uh very exciting things in this space. Um, and we're gonna trace that whole arc that they're on from uh from a GDR GDPR side project to infrastructure for the AI era, what this context really means and where it's all headed next. So I'm just really excited to have Shashanka here today. Uh with that said, let's let's bring him in. Shrashanka, welcome to the podcast. Glad you could join us.

SPEAKER_01 2:21

Oh, absolutely, Bo. Um and you said it so well. Uh you provided just the right amount of context for this audience to make sense of what we're gonna be talking about. So love how you did that.

SPEAKER_00 2:32

Excited to be here. I appreciate you saying that. Yeah, hopefully it's not too wordy, but I I um I did want to paint that picture because it's really fascinating to see um the work you guys are doing in this space. You guys are a foundational piece of this AI um revolution we're in. Um and so I want to unpack it some more. And I want to start at the beginning. I know a lot of people um might know Data Hub as this context platform, but um, of course, that's not where you guys started. It began as uh a data catalog at LinkedIn. Um maybe you can just walk me through this this journey. What problem were you originally trying to solve and and how has that mission evolved to where you are today?

SPEAKER_01 3:09

Yeah, names are so important, right? Um, data catalog, I didn't even know what that meant when I was at LinkedIn because I was just solving problems. And the first problem I was trying to solve was a problem that every data organization hits. You know, we had thousands of people, hundreds of thousands of data sets, and nobody actually really knew what anything meant or whether they could trust something and whether something even exists. So I ended up figuring out like, oh, I need to build something to solve this problem. And I ended up creating Data Hub as initially a data democracy tool. It was really about getting the right data in the right hands. Um, we used to internally think of it like Google for data. Like you go in, you type a search term, maybe you're looking for the page view data, or you're looking for recommendations data, or you're looking for jobs you may know data. And a lot of our data was, of course, in multiple different places. We had uh a big sprawling Hadoop estate, we had a lot of Kafka upstream of that, and then a bunch of operational databases. And you know, most engineers and more most PMs, when they were talking about looking for data, they're not looking for data in one specific system. They're looking for the data that they're gonna use, but they also want to trust that it came from the right place. What was upstream of that data? Oh, it's that Kafka topic. Well, I know about that one because I know the team that works on it. And so all of those implicit social trust signals, implicit technical trust signals were encoded into that first version of Data Hub. We actually called it a different name. But then GDPR really brought a lot of urgency and importance to this problem. So I used to call Data Hub a metadata platform. We still do on the on the GitHub repository. It's called it, it's called a metadata platform, not just a data catalog, because a data catalog is almost like an application that you build on top of a metadata platform. But, anyways, so we solved for data democracy. That's where we started, and then GDPR happened, or rather, GDPR was about to happen. And that importance of that problem was hey, we can't just be directionally right. We can't just be like, hey, maybe it's this data set that you probably need. We have to actually know which exact data set has PII, which specific column has an email address. When a user comes in, a member, as we call them in LinkedIn, comes into the LinkedIn website and says, please delete all my data, we have to know exactly where all of their data lives. It can't just be mostly right. And so being able to go from a mostly right system to an always right system requires a ton of work. And so we had to kind of rewrite Data Hub from the inside out, make it event-oriented, uh, and solve for that large-scale inventory problem. Where anytime a table was pushed into production or an update to a table happened in production, Data Hub would know about it within a few seconds. And so that was kind of the journey where Data Hub moved from being just a data discovery tool or a discovery-oriented tool to also additionally taking on privacy and compliance use cases. And then just before I left, and this is around 2018 to 2020, we were starting to look at machine learning as another frontier. This was pre-AI hypewave. Uh, machine learning, the classical, I'm building a model, I'm training my features, I'm shipping models to production. And AI regulation or ML regulation was coming into force, right? CCPA was a uh a regulation that was coming uh on. And we started realizing that our ML counterparts, and there was an AI team that was building ML infrastructure, had kind of stumbled onto the same problems that the data team had solved, which was I want to know if my feature is computed from the right data source. I want to have a discovery experience so I can find the right feature and make sure that it's trusted. I want to understand that I have explainability and reproducibility. And we were like, ah, seems like the same problem. And so we kind of allowed Data Hub to also support uh ML observability, ML discoverability, and ML governance within Data Hub, uh, within LinkedIn. And then we open sourced it. And so that was kind of a moment, and that was you know, 2020 when we open sourced Data Hub. And that next chapter really became all about the community. And I want to be precise here because we we didn't just throw code over the wall. Uh, we actually built the product with the community. And I left LinkedIn to create Data Hub the company in 2021, and the whole you know, first year of the company was just about building the product with the community because I realized that if I stayed at LinkedIn, I was building for LinkedIn and then kind of just helping the community, but not really able to make Data Hub work for every organization in the world, which it had the potential to. And so I stepped out of LinkedIn. LinkedIn was an investor in our first round. Um, so that was great validation from kind of the leadership team there. And then we just build a product with the tastemakers, the most demanding data teams in the world who are in the room with us, shaping it. You know, whether it's as simple as uh integrating with Tableau or something as complex as figuring out what is a data product and how should we design it, right? We were doing it all with that community. And that active engagement is really why the project kind of took off the way it did. And from there, of course, the natural arc was hey, commercial, um, take something that the community had shaped and loved, and then you know, make it genuinely enterprise grade. And that's you know what ended up being managed data hub cloud. It's the reliability, the security that large organizations require. Um, and you know, we ended up running inside companies like Block and Netflix and Apple and one of the largest banks in the world. And now we're kind of in that third chapter of that evolution, which is okay, we did kind of that commercial build-out, enterprise grade build-out of a metadata platform, but now we're moving towards AI agents and kind of harnessing the context needs that these enterprises have. And I'd really like to emphasize this point because it often gets lost. We are, you know, thinking about it and intentionally executing on it as an evolution, not as a pivot. Um, you know, you see a lot of others in the space literally walking away from data catalog and saying, well, no, no, no, that was last year. This year we are a context layer. We're actually doing kind of the opposite. We're building on everything that we're already good at. You know, the technical and operational metadata that we've already been excellent at collecting, that cross-platform visibility, and we're extending it to now cover the business side of things, you know, extending it to serve AI agents. It's the same foundation, really. And that foundation hasn't changed over time, you know, some tweaks over the years, but but it's just a new frontier with that same foundation.

SPEAKER_00 9:59

Yeah, no, that's a that's a uh that's really fascinating. Thanks for explaining that and and uh that natural sort of um that flow and progression of the company. And um I I like that that phrasing of the uh you know evolution rather than a pivot, um, because it sounds like you know, you're really just um tackling the problems that that are presented to you, and you're not really you're not concerned about fitting in with some of these like buzzwords that come and go. Um you're just working on the the foundation um and uh making an impact there. Um it's really interesting. Um so uh context is you know, we hear the context layer is this big buzzword. We hear uh, you know, vibe coding was a thing. Now I hear loop engineering. There's all these different uh the terms that come and go. But um context layer is fascinating. I want to dig into that some more. When you when you say context, what do you actually mean by that? I want to start there.

SPEAKER_01 10:53

Yeah, I would say that AI has gifted many things to the world. Uh, one of the most important things it has gifted is uh a better word for metadata. If you look at 2021, data hubs roadmap, um, and you look at, hey, what did we say we wanted to integrate into the metadata graph? Uh, we had metrics uh and we had Slack conversations and we had confluence and all of these other unstructured sources of information in there. Right then, this is before AI or anything else, because humans have always needed context to make sense of what it is they were looking at. And the the word has just registered so much better, and it has been a word that everyone kind of understands, but let's make it concrete, right? The simplest way in which I think about it is that we've always been a metadata company. We know the technical and operational state of the data. That's something a lot of people understand when you say metadata. Like, where does the data live? How is it structured? Um, how does it flow, lineage, whether it's healthy, the data quality signals. And context, and it's easier to explain it this way, is that context is what you get when you take that same metadata but you add kind of the semantic layer on top. What does the data actually mean beyond what the data says it means? Right. And if you put those two things together, then you have something that an AI agent can reason with. So if you ask me what makes context good, because you know, stale or incomplete context is probably worse than none, I'd boil it down to you know maybe three things. The first is it has to be complete. And so the breadth, I mean, if you think data is fragmented, you know, context is like fragmented a thousand more times across structured sources of information, your tables and your code, as well as your unstructured sources, whether it's confluence or notion or just PDFs lying around in SharePoint, plus the semantic sources, right? The business definitions and the metrics. Some of which, if you're in a modeling-oriented world, could be defined in code and registered as code. But if you're in a slightly less modeled world, those are implicit and in people's heads or also in unstructured space. And one technical point that I care about in all of this is that we represent context. So this multi-layered technical, operational, and then semantic and business context as a graph. And so the entities are nodes and typed relationships are edges. And so the context that an agent receives is an output of that graph that it can reason about. It's not just a pile of documents that you're just doing retrieval over. So we've kind of obviously we have capabilities for doing classic rag, uh, but really the distinction and the quality that we see when an agent is able to not just uh have semantic search capabilities, but also navigate relationships is uh you know huge for accuracy because that makes it a lot more predictable. So completeness is kind of one thing that's very important about context, both depth as well as breadth. Um, but the second thing is it has to be trusted. Just because you connect it to everything and you put everything together and you created a graph doesn't mean that it's accurate. The worst thing you want is a context layer that makes your agent hallucinate more. So um I think there are many experiments that people have done where they've kind of used LLMs to generate docs and then they've used LLMs. Like if you use LLMs all the way up and then you finally point an NLM at LLM generated stuff, you just get a ton of garbage. Um I mean, anthropic makes a lot of money, but uh I don't think the agent gets better. So the the second thing is that it has to be trusted, and humans are the experts here, right? So it has to be expert-validated, it has to be kept fresh as things change and also fully versioned because otherwise you can't reason about what changed and when. When you ask an agent, you know, you know, that question that you started the conversation with, uh, it was a revenue question or something. But usually it's a time-dimensional question around revenue, right? What was our revenue last year? What's our revenue this year? The agent has to know that our revenue calculation logic actually changed from last year versus this year. So you need kind of a clean audit history on the context graph, and it has to be trusted and validated. And of course, third, it has to be shared. This comes as a surprise to most people actually, because they're so used to working in I have an agent and I'm trying to get it to production, and I am doing everything in my power to make that agent be really high quality. And that's essentially the equivalent of localized optimal execution. So, what typically teams do when they do when they have a problem like that is they'll kind of go around and figure out what data does this agent need to be correct? What context does it need to be correct? Uh, what sort of memory store do I need for it to record whatever it's doing? So you talked about loop engineering, great, uh, great reference there. And so they kind of build a little stack, and of course, you can call it context engineering. So you build a context engineering driven stack and they ship that agent into production, and then they kind of do the hill climbing of keeping that agent working and keeping it um curated and working well. But turns out a team right next to them, who they don't know about, is building a similar agent or maybe a different agent, but it has some similar needs. It might need some of the same operational context that this particular agent is producing and delivering. And so this shared context layer and the need for a shared context layer is something that you only realize once teams start talking with each other and realizing how much redundancy and um mistakes they are kind of making in parallel together. And that is a big one that we're seeing enterprises realizing after their first round of success, comma, failure with AI agents. It's almost like you get your first few successful agents. Then the moment someone tries to talk to both agents and gets different answers on kind of a similar question, you start realizing wait a minute, these two agents are not sitting on top of the same knowledge. So we act as basically that shared infrastructure layer where you build context once for the organization. And every agent, every framework can get it through open standards like MCP. And it can stay private. You know, agents get their own space so they can store context for themselves that doesn't have to be shared with other uh agents. But there's also a shared context layer that everyone can dip into. And that's really the same idea that I laid out. Um, what is it, last year in context 2025? That was context is the annual conference that we do. Uh, and I basically said, you know, context engineering is a term that we are all talking about right now. Fast forward one year, we'll all be talking about context management. Um, and how do you do this at scale? How do you do it across all projects instead of just one off per agent? So, really complete, shared, trusted context delivered as infrastructure for AI agents at enterprise scale. That's really the role that we play in the AI stack.

SPEAKER_00 18:30

Well, and it's a it's such a foundational role. I mean, uh, like clearly what you guys are doing is is is like the um is the main pillar for all these AI companies and and more or more traditional companies utilizing AI tools and services. So I uh it's really fascinating to learn the behind the scenes uh, you know, uh details of building that context layer and what goes into it. Um I as a kind of a conceptual big picture guy, I'm curious like if maybe you can help paint a picture of where Data Hub actually sits in the AI stack. Like if I'm picturing sort of the flow of like where my data sits in like Snowflake or uh Databricks on like one end, and then you have an AI agent answering the questions on the other. I'm just curious, like, you know, where exactly does Data Hub slot in between the two?

SPEAKER_01 19:17

I would actually even question that first picture that you drew verbally, which is you've got Databricks or Snowflake and you've got an agent that's trying to talk to it. In reality, your agent is trying to talk to all the data in the enterprise. It just so happens that an important subset of your data is in Snowflake or Databricks. And so what we're finding a lot more is AI doesn't want to restrict itself to just talk to your data warehouse. It wants to talk to your data. And your data lives in many places. The data warehouse is one important place where it lives, but you also have APIs. A lot of our customers have, you know, APIs that answer questions like what's the ticker price right now? And if an AI agent wants to answer that question, I don't think you want to wait for all of that data to get detailed out into Snowflake or Databricks for it to then query the data. So what we want to enable, and this is why that stack is so important, because when we think about the stack, the lower half of the stack is your entire enterprise data estate. Could include your operational databases, it includes your streaming data, it could include your microservice APIs. And then obviously includes your warehouse, could include your SaaS tools as well. And we sit on top of that as a context layer that's kind of spidering in. Information is getting pulled in from a lot of these different sources, your unstructured data sources like Notion, Confluence, PDFs, et cetera, as well as your structured sources, both operational and analytical, and your semantic models and semantic layers and things like that. And we provide that context graph that connects all of these dots together, helps an AI agent understand that the customer object lives in Snowflake, but there's also a customer API that's online, and you can kind of reason across which one is the right one for you to answer. And then on top of that lives obviously the agents, and then on the side is the model. So, in the, you know, if you think about it, the life of a query, you know, someone asks a question to an agent, what happens next? Essentially, you ask a question to the agent, the agent turns around and says, okay, this looks like a natural language question. And I need to first understand what is the context that I need to answer this question correctly. That's when it turns around and essentially calls Data Hub as a context layer with a specific set of sub questions. Those questions could be natural language questions, or those could be more targeted MCP API questions, like we have a semantic search API, we have a lineage walk API, we have a relationships walk API, we have a generate SQL context API. There's a bunch of these APIs that are optimized for the intent that the analytics agent has. And the analytics agent will most likely call the generate SQL context interface because it's trying to generate SQL context to essentially plan its query against the database. So that's the first call that it usually makes. And then Data Hub internally walks through the graph, figures out what's the right level of context that it needs to return back to the agent for this particular question. And then the agent takes that context and then goes over to Databricks or Snowflake or the on-prem data warehouse, or could be an API call against their microservice. And in those scenarios, it could either talk to Databricks or Snowflake directly using their data API, or it could talk to the agent that lives within those interfaces, could be Cortex or could be Genie. And so you basically, if your analytics agent is living outside one of those systems, then that's kind of how it works. It basically talks to Data Hub first and then it turns around and talks to those warehouses, either using data APIs or slightly sometimes even more agentic APIs. But the real point there is that the context layer is consulted first to kind of plan your angle of attack and kind of the kind of shape of what you want to do. And then you talk to the data layer. If you flip it around and you talk to the agent inside one of those platforms, like Snowflake or Databricks, we actually support bi-directional integration. So those agents can actually talk out to us as well. And so, for example, um, some of our customers are essentially teaching Cortex to call Data Hub first before it dives in and answers the questions off of their Snowflake warehouse. That's exactly what Miro did, for example, right? We can talk about that later.

SPEAKER_00 23:48

Well, and then from that standpoint, that really just uh hammers home the kind of the trust aspect. I mean, if you're going directly to the source on the other way around, yeah, I feel like you get more accurate, reliable, in-depth information.

SPEAKER_01 24:00

Yeah, and most organizations are really not pursuing a single agent strategy. Uh, most of our customers seem to have, you know, at least two, sometimes three agents that they use to talk to their data. And one of them happens to be a platform provided agent, like uh a genie or a Cortex. But there's always like Claude in the mix. Sometimes, like, hey, my business user is on Claude, they just want to talk directly. Um, and there's also sometimes, uh, oftentimes data platform teams are just building up internal agents as their single glass pane for talking to data. And that agent is essentially the agent that they want to stand behind as a team. Because the data team often is the one that owns the interface for the rest of the organization. And they want to be in control. So sometimes they feel comfortable and confident enough to stand up their own analytics agent. It's usually like a Langgraph or a lang chain-based agent, and they are then um turning around and you know, talking to Data Hub as a context layer and then talking to the other systems as a data layer. In fact, we open sourced our own uh analytics agent just to teach the community how to build some of these agents. I mean, we're starting to see some of uh our open source community members kind of taking it and running with it. But it's been interesting to kind of seed that moment and allow people to um kind of build their own journey because now you know it's a very builder-first mindset, and people are like, hey, I could, I could build this. I want to have control over the experience, I want to have control over my context. And uh we want to enable uh customers to build alongside us.

SPEAKER_00 25:39

Yeah, uh yeah, kudos for open sourcing that project there. That's that's fascinating. I love just hearing the um you know the community that's involved and and um making these these decisions uh for for the the benefit of you know community. That's what it's all about. So they can kind of tinker with things and and build their own um you know projects and whatnot. And um and then A plus with the illustration of how Data Hub actually, you know, its role in this sort of AI stack there. Um it's really uh interesting to picture. My my um visual brain really appreciates the uh meticulous like uh illustration of of how everything all works. So I appreciate that. Um my mind was also going to, you know, obviously uh everyone's an AI company, the AI industry is is hyper competitive nowadays. Um and so I've got the kind of the competitor uh question for you. Um because obviously you're not the only ones making kind of noise about this and and and working on this context layer. I mentioned Snowflake, I mentioned Databricks, a lot of these big names have have made their own context announcements, right? Um so so what's you know, without getting too redundant, I would say, but uh I don't think you've specifically addressed this. What's Data Hub's really unique take here when a customer picks you over one of the other options uh out there? What's driving that decision, if that makes sense?

SPEAKER_01 27:03

Yeah, and I think it's a it's an absolutely um opportune moment to have this conversation because every platform literally announced context at their summit this year, right? Snowflake, Databricks, um, the cloud providers. And honestly, that's really validating because it means we are right about where this is going, we are right about how important this problem is. And the question that we have seen customers ask themselves. And then when they really introspect on that question, they decide to go with us is like whose data and whose agents are we talking about? Because when Snowflake builds a context layer, it is a context layer for Snowflake. Sure, it might be exposed over MCP, but really it ends up being a context layer that's in Snowflake. Same for data breaks. You know, they have a genie ontology layer that's kind of behind the genie agent. So it's like you don't really get to the ontology layer, you get to Gini, and then Gini does its work to get to the ontology layer and give you better answers. But it really doesn't allow you to, you know, own your context and own your own business context for yourself. And same for the for the cloud vendors too. And the reality is no real enterprise lives in one platform. And a lot of customers nowadays, and we saw that with iceberg as well, they really don't want any lock-in when it comes to um any proprietary aspects about their business. And we saw that with data, uh obviously, and open compute and you know, data storage became a thing. And uh iceberg, I think, had its moment because of that. And so most of the enterprises nowadays want to stay future-proof, both when it comes to their model provider as well as when it comes to their data provider. And so we see that the same thing will happen with context, and we we we know that it's already happening because most of our customers seem to have like data breaks, but they may have one team already playing around with Snowflake. Um, or even if they are all in on one warehouse, they might have a dozen other SaaS sources or operational plane data that they want to bridge into AI. So a context layer that only sees one vendor's world is by definition always going to be incomplete unless an enterprise fully, fully, fully standardizes on that one vendor as for all of their needs. And I think there will be companies like that. And those are not the companies we want to be applicable for, right? If you're if if you're fully shrink-wrapped taking everything from one vendor, including your entire data stack, then you should probably just take everything from that one vendor.

SPEAKER_00 29:41

Would that be my my mind goes to like a company like Apple, who's really uh takes a privacy first standpoint? Is that something like a use case would become a privacy standpoint or uh something else?

SPEAKER_01 29:52

Well, our unique take, and it's the whole game really is that we are vendor neutral by design. Uh we sit across all of it. Um, we don't have a compute engine that's trying to bump up credits uh uh on you, right? It's like that's the data layer's job. And so our game is we are vendor neutral by design, complete context, and I mean complete deliberately across everything that you have, across structured, unstructured, across every source. It's expert validated, it's fresh, and it's yours, right? That's the most important thing. Uh, it's yours to integrate with. You can push to it, you can pull from it, you can subscribe to it. Uh, it's not hidden behind another agent that you have to uh you know get from us. And so while there's uh a lot of the vendors and uh completely understand those incentives, right? Their incentive is like lock in and kind of uh bring people in and keep them there and pull as much context in and uh make it uh hard for them to leave the platform. Ours is kind of the opposite, it's interoperability, giving our customers the choice to make future decisions that don't require huge migration efforts. Right. If I have chosen Databricks this year, and maybe in two years I want to say, hey, maybe Snowflake has actually leapfrogged them and they've done this new cool thing that I want to try out. Oh, I can't move because now my context layer is fully locked into databricks and it's a you know a crazy idea to even move. So our really uh offering is about future-proofing a lot of your data and AI stack decisions. And this is true for models as well, right? So the you know, you can see models and data and then the context layer, right? These are kind of the three pillars, and we think that the context layer should be interoperable and be independently chosen to interoperate with the data and the model layers. Of course, underneath that, there's a lot of engineering. So the engineering part is hard work done over decades of investment. We were built event-oriented from day one for massive scale. Um, that's the reason BlackRock runs us, right? Behind kind of the big, um, their big Aladdin suite. I mean, they were talking, we we talked about it publicly at uh Snowflake Summit. And a lot of that semantic knowledge is derived from real query patterns, um, and it's expert-validated. We've got that loop going where naturally people think about Data Hub as that central context layer where you come in, you validate your context, and you go away. But I would say engineering is engineering. I mean, we are proud of out our engineering. Um, it ends up uh winning us most deals when people really put scale and interoperability and uh the kind of demands that AI truly places on a context platform, you know, I think we're we're fine there. Uh we do really well. But the real, the real positioning and the real reason why people should buy us is about architecture and where things live, like that reference tag that I described. So being able to say you should have a context layer that's yours to operate, that's future-proof, that's built on open interoperable standards, and not something that locks you into a particular data decision. That's really why customers choose us. And frankly, why uh a neutral player should be win. You know, we see that with integration layers all the time. Yeah, neutral players tend to win.

SPEAKER_00 33:14

That's uh yeah, that's a great answer. I think it's from what I'm hearing, it's sort of like you're the you're the the Switzerland of of this uh you know, context data problem, and and what you guys are doing to um to work uh your work in this field. Um you do it really well. Like you mentioned, the cross-platform, no lock-ins, uh, works with sort of whatever agent you've already got. Um, I think that's a that's a strong position to to hold, right? Um yeah, so that okay, and you mentioned this this forward-looking roadmap. Obviously, like you've um you had some of the that validation seeing sort of like uh some of these uh established or you know, other companies, Snowflake Databricks, uh like release their own sort of uh have their own sort of contextual announcements. Um that's just validation that you've seen the writing on the walls years ago. Uh and so let's look into this this upcoming, you know, this product roadmap. What's what's coming down the horizon? Uh like you've I I know you just shipped a big release aimed at analytics agents, um, and you're going a little bit deeper on that. Are you expanding into other kinds of use cases? Just curious to hear what's next on that front.

SPEAKER_01 34:23

Sure. Um, yeah, just this summer we announced the context platform v1, and this was like a labor of love with a lot of our customers who were our design partners during that exercise. So we're now in kind of officially private beta mode, onboarding more customers, and then soon we'll be in public beta and then GA with that offering. Um Miro, who was a customer, uh, came out publicly and provided a case study for kind of how they were able to achieve um high analytics agents' accuracy with uh the Data Hub Context platform. You know, their Snowflake Cortex agent was doing about 50% accuracy, which is like a coin flip. And then with Data Hub Context, um, they were able to get to uh almost 90%. So that was a the really great study, uh really great uh outcome. And of course, it is not just the technology on its own, it's about how the technology enables humans to add that additional context layer, uh context curation on top to really take that, you know, we give that first step a function of context quality, and then we bring in the humans and they can then take it up to that 90 plus percent level. So that is great validation, and we're we're gonna see a lot more case studies like that come out uh over the summer. Uh but in terms of where we're headed, uh I think I'll I'll talk about it in maybe two ways. Uh the first is uh by use case, and the other is uh perhaps what the role we play. On use cases, obviously the the center of gravity is analytics agents, talk to your data. You know, the near-term roadmap is mostly about adding depth and capability to what we've already announced. And then uh enabling more types of data agents. So not just talk to data, but a broader set of agents that need to operate on enterprise data. And I'll I'll be a little careful here because that's exactly where the line sits and us providing context versus providing the agent itself is, I think, an interesting question. And I think we are keeping our uh options and opinions open on that. I mean, for companies that want to get uh a bunch of de facto agents from us, I think we'll definitely explore that option. But in general, we want to enable um companies to build on top of us as much as we want, as much as they want. Um and the other dimension, so that's kind of broadening beyond analytics agents and supporting, you know, data quality agents, compliance agents, you know, that kind of stuff is definitely kind of in the roadmap and also things that our customers are already starting to do with us. Uh, the other dimension, I would say, is the role we play over time. You know, right now, and even when I described kind of that visual example of how a query um happens and how Data Hub gets consulted, I skipped over an important piece, which is what happens after the result is delivered back to the user. Right? The thing that happens before is context retrieval, and we provide that already today. And the thing that happens after that answer is delivered is context corrections and context memory. And so, really, that's the frontier that we are already on. Um, we already have in our MCP tool set and ability to record observations back into the graph. And so that agentic loop, the memory system that we're building, enables agents to carry forward persistent understanding across tasks. And that neutral shared layer that we provide is really that natural place for that to live, for the same place that it's a natural place for even that initial sourced context. So, running through all of that is a theme that I care a lot about, giving enterprises and organizations ongoing visibility and transparency into everything that's going on with their AI initiatives. The, you know, we've always been about giving you an end-to-end view of your enterprise data landscape. And now it's really about giving you an end-to-end view of your enterprise data landscape and your AI adoption journey. Like how ready is your data even for AI? Most companies don't exactly know. Um, and a lot of the AI pilots that fail are not, as you know, because the agents were bad or the model was wrong. It was more because the data wasn't even ready for it. And so the way we help people understand that is by giving them visibility and transparency into that continuously. And so being the layer that can not only be the retrieval layer, the memory layer, but also the visibility and transparency layer that's continuously telling them how's your AI adoption looking, which domains are doing well, which teams are doing well, which areas of your data stack is actually ready for AI, and you can actually start accelerating AI outcomes in that part of the business is super important for data leaders and AI leaders. So if I had to summarize, you know, more depth, obviously, in kind of the analytics uh use case, more types of agents, you know, supporting different kinds of agents beyond just analytic agents and really evolving from just context to memory and all on the same neutral infrastructure. That's that's really what our roadmap looks like.

SPEAKER_00 39:29

Yeah, that's you've got a lot to work on and exciting uh things to tackle. And yeah, just broadening that, uh deepening that analytics piece, um, but also broadening out beyond just sort of the the talk into your data side of things and giving companies a really a real sort of like beat on on where their AI adoption is actually uh um you know, what it's how it's how it's standing, you know, what is what is it, um is it currently making a big uh impact, that sort of thing. Um I'm excited to see where this this industry is going. I mean, I know it's it's really difficult to say um, you know, where things are gonna go in uh six months, a year, let alone three or five years from now. But um it's fascinating to hear. I want to switch uh I want to switch gears for the last couple of questions here to focus a little bit more on the big picture side of things, have a little bit of fun. Um and I want to first start with that one big thing question for you next. Um every founder, they they every founder have that has that one thing they wish they could shout from the the rooftop. The if only prospects knew this line, uh if you could get one message across to a potential customer with no caveats, what would that one big thing be?

SPEAKER_01 40:37

Well, we actually flew a plane um at Snowflake Summit. So uh there were people on the rooftops looking you know further up. That was that was fun to watch. It had a big banner that said uh context changes everything. Uh but if there was one thing I could shout uh from that plane, it would really be that uh Data Hub is the most widely deployed context platform in the world and it is running at the world's most significant companies. To the extent that you know metadata is context, and it is, uh nobody has more of it in more places and at more scale than we do. So that's not a roadmap promise that's in production today. I think that's probably the one thing people don't even realize.

SPEAKER_00 41:17

Yeah, I love it. All right, and then I want you to sort of take your CTO hat off and put on your fortune teller hat for the second question. Uh, looking at uh over the next, you know, five years, let's say, what's what's the emerging trend you think is going to have the biggest impact on this industry?

SPEAKER_01 41:34

Well, I think about this all the time because you know, when you start thinking about the future of the workplace, the future of um software and the future of the software supply chain, um, as well as how businesses will be constructed in the future, you realize that obviously the models will keep changing and they'll keep getting better. Um, there'll be a new best model every few months and very soon every few weeks. The platforms will keep competing on data gravity and you know, owning your data. You know, there'll be a new database getting built every few months or every few weeks that does better at a certain class of data than it previously did. Um, and my bet, and our company's bet is that there will be a winning context layer that sits between the model and the data. And that winning context layer is going to be and is going to need to be model agnostic, is going to need to be data player agnostic. It needs to be interoperable and it needs to be open. So the platforms will keep competing, um, the models will keep changing. But what enterprises need as connective tissue that doesn't change underneath them will be this layer. Um, and you know, that layer is what we are aspiring to be, and that layer is what we think is going to be kind of the durable piece that lives while a lot of the software stack kind of compresses down and up. Um, we we're basically positioning ourselves as the layer that connects data to AI and the platform that you use to build uh kind of your agents on top of, not in terms of running the agent on it, but you know, the context platform that you can integrate your agents with. And that neutral and open layer is where the durable value will be retained within the enterprise. And that's why it has to be theirs, not a vendor's.

SPEAKER_00 43:24

Model agnostic open and interoperable layer connecting AI to data. Um, I love it. That's that's a great uh mission statement and and um uh fan a great bet to have. I think uh you guys are doing all the right things, and I'm excited to see where you guys uh um go from here. And would love to have you back in in the future to to talk updates. Um uh for folks who are interested in learning more about data hub, maybe getting in contact with you and your team, where should they go? Datahub.com.

SPEAKER_01 43:55

It's the simplest uh thing to remember. Um, we've got the team there, and our open source community is on Slack. So just go to datahub.com, find our community link, and just drop in to our Slack community and uh say hi to us.

SPEAKER_00 44:10

Perfect. Datahub.com. All right. Um and uh Shrashanka, um, I I do want to uh I don't want to throw you on the spot too much, but like you're an impressive figure. You're you're extremely well spoken. Um you've worked in this industry for uh a number of years, refining your your talents and um expertise. Uh I'm just curious, like, do you have a resource that has really resonated with you to kind of shape your your your journey? Um that kind of comes to mind, maybe it's like a podcast, a book, some kind of resource in general um that you might recommend listeners to read?

SPEAKER_01 44:44

Uh of course I end up uh listening to SourceForge quite a bit. Uh, but in addition to that, um, you know, Lenny's podcasts, of course, is uh pretty top on my list of uh podcasts to listen to. I recently heard John McNeil speak about some of the operational processes that he went through in transforming Tesla and SpaceX and a bunch of other companies. Uh, I found that uh talk quite um quite foundational. In terms of how I have honed myself, I am a big believer in actually spending time with yourself. And by that I mean like spending time inside your head. Uh so I do spend quite a bit of time thinking and reflecting um just quietly on things that are uh going on in my head. And oftentimes the clarity that emerges from really sitting quietly and thinking through a concept. I'm very intentional about first thinking about the problem and coming to my conclusion about that problem before I start reading about what other people are saying about it or thinking about.

SPEAKER_00 45:55

That makes sense. It's because it at a certain point it kind of becomes too much noise and you can't really focus in on any one particular area. And um I totally, I totally see that and resonate with that. Um that's thank you for sharing that. I appreciate that. That's that's such an important takeaway. And this is that's the cherry on top and the um little bonus uh information for for listeners who are sticking sticking with us to the end here. Um so thank you. Thank you again for for humoring me and um answering that on the spot there. Um Shershankah, thank you again for all the insights. Um again, I would love to have you back and uh and talk more updates on this contextual layer we keep hearing so much about.

SPEAKER_01 46:32

So of course. Uh it's been a pleasure, Bo. And I know we we went way over uh the amount of time we were planning to speak, but as you can tell, uh I can speak about this all day. But it's been a great chat. Thank you for having me on.

SPEAKER_00 46:46

Of course. Um, thank you all for listening to the Source Forge podcast. I'm your host, Bo Hamilton. Make sure to subscribe to stay up to date with all of our upcoming B2B software related podcasts. I will talk to you in the next one.