Skip to main content

Enterprise AI at Scale: Serving Google Gemini on Databricks for Production

Summary

  • Google Gemini models are available on the Databricks Data and AI platform through AI Gateway with unified governance, rate limits, cost controls, and Unity Catalog logging — accessible via the ai_query() function for one-line inference on billions of rows.
  • HG Insights processes 7 billion documents using a progressive pipeline where rules and custom BERT models handle predictable cases, while Gemini Flash acts as an agentic enrichment layer that reasons over incomplete company profiles to resolve missing attributes.
  • The partnership delivers 2,000-plus joint customers access to an integrated stack covering models, Managed Agents with secure sandboxes, Agent Cards for discovery and reuse, and Lakeflow for orchestration — enabling 3x entity coverage growth and 10x faster detection at HG Insights.

Enterprise AI at Scale: Serving Google Gemini on Databricks for Production

Watch: Enterprise AI at Scale: Serving Google Gemini on Databricks for Production
Foundation models are moving from chatbots into enterprise data pipelines. Google Gemini and Databricks together make this real for production at massive scale: enriching records, classifying documents, resolving entities, and running agentic workflows on billions of rows. Google Cloud and Databricks deliver unified governance via AI Gateway with cost controls, rate limits, usage tracking, and Unity Catalog logging, all accessible through ai_query() and Model Serving.
At HG Insights, a production pipeline processes 7 billion documents where rules and custom BERT models handle predictable cases, and Gemini acts as an agentic enrichment layer reasoning over incomplete company profiles and resolving missing attributes. Learn how to balance cost, accuracy, and latency at scale; design progressive pipelines using cheap keyword filters, SLM models, and frontier models judiciously; use Managed Agents for secure governance; and apply the technographics pipeline pattern to extract structured signals from unstructured text. Lakeflow orchestrates the workflows, dashboards provide visibility, and human-in-the-loop validation ensures production trust.
🤝

Chapters

FAQs

How do you use Google Gemini models on Databricks?

Google Gemini models are available through the Databricks AI Gateway and can be invoked using the ai_query() function for single-line inference directly in SQL or notebook workflows on the Databricks Data and AI platform. AI Gateway applies governance controls including rate limits, cost controls, and Unity Catalog logging to all Gemini API calls.

What is Gemini Flash and why did HG Insights use it for their data pipeline?

Gemini Flash is a cost-efficient variant of the Gemini 3.5 model family designed for high-volume workloads. HG Insights chose it as the reasoning layer in their technographics pipeline because processing 7 billion documents requires a model that balances analytical capability with the economics of running at massive scale.

How does HG Insights' technographics pipeline work with Gemini?

The pipeline uses a progressive filtering strategy: cheap keyword rules eliminate non-matches first, then custom BERT models handle the predictable majority of cases, and Gemini Flash is invoked only for ambiguous or incomplete company profiles that require reasoning. This approach drove 3x coverage growth and 10x faster product install detection.

What are Managed Agents in the Gemini Enterprise Agent Platform?

Managed Agents are a governance capability that runs agent workloads in secure, isolated sandboxes with defined identity and access controls, so enterprise teams can deploy agents without exposing sensitive data or infrastructure. They are part of the integrated Gemini Enterprise Agent Platform stack that covers models, agent runtime, tool registry, and security.

Full transcript

[00:09] Hello everyone. My name is Abhishek. Uh sorry for a bit bit of the technical difficulties, but we're ready to go now. Uh nice to meet all of you and today we'll be talking more about how can we actually use Gemini in order to extract business values uh within your Databricks uh environments itself, right? I'm also
[00:25] share uh joined with Shailesh and Badri from the H3 insights team who will be sharing more on how they integrate Gemini into their enterprise workflows. All right. Uh so before we get started, right? Why and we get technical, why exactly Google
[00:42] Cloud and Databricks? A few numbers. Uh one, 2,000 plus joint customers and 65 plus uh percent plus year-on-year year-over-year growth. Uh our we have a very great first-party relationship, unified billing via Google Cloud and
[00:59] Databricks marketplace, and this is done at scale. Two, the serverless uh data intelligence stack with Databricks SQL, notebooks, and workflows are GA. Uh this is governed by the unifi- Unity catalog and natively integrated with
[01:15] BigQuery, Looker, uh and our agent platform. Uh that native integration is what makes today's demos and content possible. Uh and of course, uh Databricks is a Google Cloud technology partner of the year for 2 years running.
[01:31] So, let's look at a whole zoomed-out big picture itself, right? So, this is the entire picture with Gemini Enterprise Agent Platform. So, everything starting from your models at the bottom to all the capabilities that are required to
[01:47] run agents, to build them, scale them, and govern them. Uh and to the apps that you or your partners build, right? I won't walk through each and every box, but think of it as a map. The takeaway is that it is one integrated
[02:02] stack. Uh models, agent runtime, a registry for agents, tools, skills, uh along with I agent identity and security, which are the most crucial pieces. Uh so, you're not stitching together five different products in order to
[02:18] govern and scale your agent. You actually have everything available in one single integrated stack. So, today I'll be going deeper and showing you more about two specific things. One is the models, and the second is the managed agents. And governance, of course, runs
[02:34] throughout the platform. So, let's start at the foundation, which are the models. Everything starts with the models. We recently launched the Gemini 3.5 family of models, which come in two variants, which are the Flash and the Pro. Uh Gemini 3.5 models are great at
[02:51] advanced coding and reasoning, uh as well as tool and workflow execution. Uh these are built specifically to power uh agentic uh operations. Remember this one, because our managed agent demo later on is going to use this uh Gemini 3.5 Flash model.
[03:09] Flash, which is GA, excels at coding, reasoning, and real-world tasks. It beats 3.1 Pro at higher speed uh for 1/3 of the cost. Your This is your workhorse model for high-volume on data work.
[03:25] Uh we have 3.5 Pro, which is going to be coming soon, uh which excels at coding uh and uh is our most intelligent multimodal model. This has been tuned specifically for agentic coding and layered and multi-step patterns. Uh Our practical
[03:42] guidance and recommendations for your use cases here would be to use Flash for anything to do with scale and most agentic work. And when you want that maximum reasoning, that's when you can switch over to the pro models. And everything I show you today will run on
[03:58] flash. And the reason why flash matters so much to this room is specifically because of the cost. So, let's take a look. Uh this slide should essentially change how you think about cost, right? Uh Y axis is the intelligence, and Gemini flash
[04:15] pushes the intelligence up while pushing the costs down. Uh it sits in the corner that you want. Uh specifically on terminal bench 2.0, which is a benchmark for real-world agentic operations that include multi-step terminal tasks like
[04:31] compiling, debugging, and so on. Flash is right at the frontier at a fraction of the per-token cost of the compatible frontier models. Why do I keep hovering about cost, right? In a minute, I'm going to run this model on
[04:46] Databricks uh live on Databricks data, which runs it row by row. And think about it, when you look at billions of rows at 3x the cost, this is not a rounding error, right? The cost compound up. So, you ideally want to have costs minimized, and this
[05:04] essentially makes the difference to what and when you can run on production. So, we now have a fast, cheap frontier model, and let's actually point it to your Databricks data. So, the very first thing that you essentially would do is configure your
[05:21] AI gateway. So, I'm just going to exit out of my screen here for a second. Uh and this is the Unity AI catalog with the gateway. And as you can see, Databricks Gemini 3.5 flash model should already be configured here. You can set up usage tracking, you can
[05:38] set up inference tables to actually track what kind of you know requests and responses can be logged. You can also set up rate limits in order to ensure that unified governance and ensure fairness to all your users along
[05:53] with various guardrails, right? But once you have actually configured that, that's when it gets really, really easy. Uh because once you have that, you go to Databricks query and then you can just run the AI query, which is here, uh with the
[06:08] Databricks Gemini 3.5 Flash endpoint, and you can provide a natural language prompt to all the rows in your table. And when you do it, for example, here I say in one sentence, "Assess the health of this account, right?" So, in this scenario, I'll I'm essentially demoing a
[06:24] fictitious company called Symbol, which has a product that uh we're selling, and then we're essentially looking at our customer accounts that we want to take a look at. And you can see that Gemini Flash actually goes through all the rows, does the joins for me, and then is
[06:40] able to give me a summary of exactly what is happening within that account, right? So, it's as simple as calling your AI.query SQL operator along with the prompt, and you're good to go, right? Uh super simple and super easy to use. So, let's go back into the presentation. All
[06:58] right. Uh yeah, so step one was to essentially configure the gateway. Step two was to uh use AI.query in order to fetch the actual data from your uh uh and actually run the inference on
[07:13] your data itself. Okay, that's awesome. But once we have the data, right? Uh and once we are past the one-step workflow, we actually want to stake certain actions on this data. And that's where agent essentially comes
[07:30] into play, right? And I'm sure all of you are very familiar with agentic operations. Uh but I specifically want to call out one sentence in the slide, right? Which is "Technology changes exponentially. Organizations change logarithmically, right? So, what does
[07:45] this mean, right? So, we are seeing code being developed at in at a rapid pace. Like, code is being Code is now cheap. We are able to generate as much code through various coding agents and build a lot of applications and agents and whatnot, right?
[08:02] Uh but, what is missing essentially is the pace of your organization in order to scale and govern these sort of uh systems, right? When it comes to security, when it comes to governance. And this gap is where most enterprise enterprise AI use cases actually stop,
[08:19] which is from the gap from building a prototype to actually taking it a scale, right? So, what we're trying to do here is actually dramatically accelerate the timelines to take your agents to production. So, how do we actually do
[08:34] that? Uh All right. So, we have a unified architecture with Gemini Enterprise Agent platform that uh this slide actually talks about. So, if you take a look, when the user query comes in through the front end, it actually goes through the agent gateway,
[08:51] right? So, the agent gateway essentially acts as the uh front gate in order to essentially understand what are the different models or what are the different agents that are coming in, what the according to the user query, what agent should be invoked, what tools
[09:07] should be called. And that's where each agent has a unique identity governed by agent identity. And this is essentially what provides the agent access and authorization for various tools on what it should do and what it should not do, right? And of course,
[09:23] along with agents, you just cannot have agents running. We have various uh things like skills and tools that are actually equally important in order to increase the quality of your workflow. And that's where all of this is automatically tracked within your agent registry uh as a single unified place
[09:41] for all all artifacts generated by agents as well as agents themselves. And of course, as you can see AI observability and governance goes through the entire stack and not just at one point of the stack, right? So, today we're actually excited to
[09:59] share a a very interesting product that we launched at Google IO, which is called managed agents on agent platform, right? So, think of this as you have an idea for an agent, you prompt that agent to Gemini. Gemini actually builds the agent
[10:16] and runs it for you, right? So, you don't have to actually worry about provisioning a sandbox, figuring out security, figuring out what tools to call. All of that is essentially packaged by Gemini as part of our anti-gravity harness, and what you get is a
[10:33] specialized agent that actually is able to use our advanced coding harness for anti-gravity within a secure safe sandbox, and that entire infra is actually managed by Google. So, how do you think of this, right? So, we have two parts of the API. So, the
[10:50] very first API is the agents API, which essentially acts as the setup. So, this is a one-time call where someone like an administrator essentially sets up this agent. And this actually is responsible for the actual governance. So, you you
[11:05] can imagine that you write your agent configuration as like a JSON file, and we'll take a look at an example next. And you call that agents API with that particular JSON file, and then it stores that configuration in the agent
[11:21] registry. So, what this allows you to do is apply governance at a unified place, and this agent is then immutable, right? You can then only one person can edit it and then that updates are global. So, no other person can
[11:37] actually go and edit the agent while it's running. And when you actually want to use the agent, we actually offered another API called the Interactions API, which is actually responsible for provisioning the sandbox from this stored agent, actually applies the
[11:53] intelligence, attaches the tools that you have defined in your configuration, and then makes the API call. So, this is how it looks like in a nutshell, right? So, think of it as the the Agents API being your control plane for building an agent, and then
[12:09] Interactions API as the data plane for building the agent. So, you call the control plane once in our the Managed Agents API once in order to set up the agent, and then you can call the Interaction API any number of times in order to actually interact with the
[12:25] agent that you just created, right? So, how does this actually apply to Databricks? So, this is an example of a business intelligence or a usage product usage agent that I built, which is connected to the Databricks SQL MCP
[12:43] remote MCP that was created by Databricks, right? So, you can see the configuration for building this agent is very simple. You have the agent name, you have a display name for the agent, and then you have a system instruction, right? Similar to how you would govern all agents, you essentially give it a
[12:59] system instruction to steer it and ensure that its quality is great. And along with that, you define a list of tools that you're passing to it. So, in this case, I'm saying that, "Hey, I have an MCP server by Databricks, which is my and it is at my URL, which is provided
[13:17] in the slide." And then you of course have to attach your auth headers, which is again automatically managed. And then, once you call the Agents API, you can actually use this using the interactions API. So, we in fact took this one step further and when you
[13:32] actually interact when you actually add this to the when you actually make the agent CPI call, you are actually able to see this as part of your uh enterprise itself here. So, if I come here, you can actually let me just zoom
[13:49] in for a second. Yeah, so you can see the agent name uh So, this is essentially the agent card, right? We are advertising the agent cards in order to ensure that other clients that we may have, like for example, Gemini Enterprise, can actually
[14:06] discover our agents and be like, "Hey, I see that there's a symbol usage intelligence agent. Let me call this agent for my particular query that I have which needs to do some product intelligence, right?" So, you can see the agent card that is defined. You can
[14:21] also see like what are the capabilities that it has, what are the skills that it is attached to, and so on. Uh and how do we actually use this, right? So, if we go to, for example, Gemini Enterprise, which is a client, you can actually go there and just add
[14:38] the agent and you can see symbol usage intelligence and then ask your particular question. For example, it could be like, "Where should we focus expansion? Show the accounts whose active users exceed their license seats and rank this rank by seat intelligence,
[14:53] right?" Uh and then I'm just going to fire off the query quickly. Uh and yeah, you can see that it's starting. You can see the traces coming up, which is sandbox attached, which means that we provisioned a sandbox in less than 2 seconds and then actually
[15:08] have started running your actual query, right? So, we actually look at all the tools that the agent has been provided, all the skills that we attached to it, and all of it is running in a safe, secure sandbox that is managed and provisioned by Google infra. So, while this may take a while,
[15:26] so in the interest of time, I'll actually open a chat that I did on this, right? So, for example, it will give me back an answer saying that, "Hey, we should focus expansion on these three specific accounts, right?" Which is like Coral Media, which has a
[15:42] utilization, and then it says the source where it got it from, and then provides me all the accounts and whatnot. And along with that, it also shows me what specific query from Databricks it actually ran. And not just that, with Gemini
[15:57] Enterprise, you can actually connect it to what we call as A2UI, which is agent to user interface, and actually build very, very interesting dashboards on the fly like this, right? So, this is completely 100% generated by the antigravity harness,
[16:14] which actually queried Databricks data, got the data, analyzed it, created a custom dashboard, created custom reports and charts for you, and renders it in line. So, you have intelligence at the edge wherever you are, right?
[16:30] So, yeah, that's how this entire workflow runs in a nutshell, and that was a short demo on how this works with Gemini Enterprise, and how do you actually create and manage your agents as well as
[16:46] govern them. I'll now pass it on to Shailesh to share more on the workflows around ATInsights. Thank you. Uh Go Gemini.
[17:05] I attended few sessions in this conference, and one pattern stood out to me. The teams who are trying to deploy the AI solutions at scale in production, actually, not just a POC. They're trying to balance three parameters. Uh first one is uh
[17:23] of course, the cost. Like, cost of token is at the scale, it matters most. Then, uh you can call accuracy, like that's where your e-vals and harnesses come in play. And the third is, of course, the latency and speed. At Edge Inside, we also face the same
[17:39] problems uh when we tried to deploy our conventional ETL pipelines and made it AI-enabled in some way, and here we are talking about um billions of document processed per week. So, that means every small model change or every small amount of tokens which we
[17:55] utilize reflects into hundred thousands of dollars. Um to give an example, oh, sorry. I'm here. To Since this is a AI and data conference, I'm not sure how many of you will be familiar with the GTM or sales
[18:13] concepts. Just to give an example, just imagine a scenario where uh there's a VP marketing of an B2B SaaS vendor, and and that person is trying to ask a simple question, "Where is our top competitor's most vulnerable to displacement in our target
[18:30] accounts?" Three things, top competitors, displacement, target accounts. So, depending upon what that SaaS vendor does or their ideal customer profile is, with Edge data, we can answer these questions. Yes, in your account, there
[18:46] exists a company somewhere, uh uh and they have a marketing department in London which has around 200 users or employees. They installed a competitor products like 3 years back, and it's up the contract is for renewal.
[19:01] And um their usage patterns is signal is low. Uh account and they're also researching for the competitor platforms, right? And um their IT spend is year over year 15% up, and so and so forth.
[19:17] So, this is this is all for answering this question, right? We have to this is done from our multiple pipelines working together. And, we do this for every company, every product, every intent signal, every contract. We try to scrape, we try to
[19:33] buy the data from various places, and we curate this information in one place. And, we've been doing it for last 10 12 years. We have a history of data and analysis done for last around 28 million companies now, and
[19:50] around 60 odd K products. So, in short, Edgy Insights is a data and AI intelligence revenue growth company. Um, and it provides the signals for any GTM strategy team with this amount of
[20:08] data like firmographics is very talks about what are the different types of companies, what are the departments, how many how many employees are working in that department, what sort of software they're using. Uh, spend such estimates, how much they're spending on particular category of software. Uh, we also have buyer
[20:24] resource signals, uh, competitive footprints, a lot of information which enables your GTM teams to create various data back strategies like market sizing or account prioritization, white space analysis, renewal risk, and so and so forth.
[20:45] So, we started this ETL AI pipelines project like 6 months back. And, we decided few tenets for our platform or or for the pipelines. And, one of those was do not pass everything to LLM at first place. Try to do and solve the problem deterministically.
[21:00] And, only and only when you have a reasoning needed, go for the frontier model. Over a period, we have created a huge domain expertise and curated large amount of data with our um expanded team of labelers. And we have uh we don't want to break
[21:17] our data domain knowledge biased with the LLM. So, that is very important. Um the approach we took was to build bespoke models. Every model is just doing a very small task. And and that means we can evaluate the
[21:34] results or validate the results very specifically rather than one big model doing everything for us. So, we have trained multiple models based on our data. That is and that because of that we are able to validate the data or the results of the model very accurately.
[21:50] And of course, um for entire pipeline, we had to reimagine our human in the loop or validation process. Everywhere the harness which we built to validate the results so that there's no hallucination in the data. Um so, that's what
[22:05] In fact, that was the most time we spent on this particular project rather than building a pipeline. So, in short, uh Oh, sorry. My bad. I think uh something missed.
[22:26] Oh, sorry. I think I'm missing few slides. To give a Okay. Um Okay. So, pipeline which I'm going to talk was going to talk about was the Technographics pipeline.
[22:42] And the the crux of the pipeline was we we ingest um lot of our data from, let's say, Open Crawl, which is a free internet dump of internet every week or every month. We also buy data from vendors. We also
[22:58] curate the data. And here we talk about like billions of documents, unstructured text documents. It can be resumes, job descriptions, use cases, job posting, anything in the world. Once the data lands, um we have a data acquisition team which
[23:14] sanitizes the data for PII, and then the clean data lands on our data bucket. Once that thing happens, the first step is the data prep where all the junk separations happens based on the metadata. For example, we get a resumes right from generator to CRM.
[23:31] Right? So, we we try to make sure based on our models, and because this is a technical product extraction pipeline, we try to extract only those products which are our resumes or job posting related to our domain. Once that happens, then the series of SLM extraction starts for extracting a
[23:49] specific signal from the document such that it can be validated. Once the new signal comes, we vectorize it for our ANN searches so that all our deterministic or fuzzy logic searches
[24:05] come into play where again they are validated for a confidence level. Once it happens, we get a low or medium a confidence a signals. Those signals again go back This is where we use actually the frontier LLMs to make the decisions or
[24:22] our reasoning against the signals. And then reconcile and emit. For the new products or for which are found basically in a document, we then go for a human in loop and the deep enrichment process. Uh when I say product, what is a
[24:37] product? Sorry, I I'm missing couple of slides. That's That's reason it might be a little bit difficult to understand. A product for us is anything, let's say a company is using Databricks. Now, we would like to know why is this company using Databricks? When did they install
[24:54] it? Is there any contract signed about like for this company from the data bricks and so on and so forth. So, for example, this product Cortex, right, from Palo Alto, we we can detect, okay, this is the this is what the product does, edgy category, Gartner magic quadrant, use cases, how
[25:11] is it they are doing a billing, and and how is the deployment model, what are the family and all those other things. Why is all this thing important? This is important because we would like to sell to the people who are selling the software, right? So, this and then we
[25:28] also detect how many of how many companies in your account has installed Cortex or any replacement. And because of this pipeline, we increased our product catalog by 3x, which is around uh 100k products, and 10x was a
[25:46] detection time reduction. Before AI or this pipeline, most of our curation was handcrafted by the labelers. And that means one product used to take a week, basically, to curate. With this new tech, we are able to exponentially increase the products
[26:04] and also reduce the time. You'll find that we actually this is a cost reduction, also, because as I said, it was a rule-based engine. Every time a rule changes, we had to pass the document again from the scanner. With this new pipeline, we are uh
[26:19] passing a document only once in its life for a product extraction. And of course, we have product more product resource stuff. Yeah, I think that's that's the my part of presentation, but for interesting part, how this actually works in the
[26:34] demo, that will be more exciting than this boring slides. Uh I'll hand it over to Badri.
[26:53] Are you able to hear me? Thank you. Okay. So, yeah, we So, whatever Shailesh presented, let's see that in action. And what we do is like we try to bring the data to wherever your agent is. For example, you
[27:10] could be using cloud code. I'm using cloud code here. We have an MCP server that with all our data presented it. And it can help you analyze your competition, right? Like let me ask it some question here. Can you
[27:25] show all the tools available from HG Insights? Make a summary of it, right? So, So, this is our HG Insights, yeah. Probably I misspelled it. But it goes
[27:42] and finds out all the connects to our MCP server. It gets all the company intelligence, contacts, products, and vendors, whatever we have to provide, right? So, for this conference, I could be asking things like okay, um
[27:57] who how is the market share of Databricks compared to Snowflake, right? Like their biggest competition.
[28:12] In terms of in in the last uh 4 years, let's say. Because our one of our bigger other big motives like we we have a lot of time series data, right? We've been curating this for 10 years now. So, let's keep doing this. But that's the
[28:28] as it's doing it, I have done this before. And I think we have like some more time. So, I'll show you some analysis that I did. I just asked Cloud to create this as a report so that I can present it here. So, if you look at this data, it's very interesting. Snowflake was way ahead
[28:44] like 4 years back in 2022. And you can see that Databricks has been catching up. This is like every distinct company that is using Databricks. And this is like very high-level data, right? But we specialize in going into finer grain details. Like we can tell
[29:01] you like exactly in which market is this happening, right? Like in Yeah, in which in America or Europe or Asia Pacific, we could give you that fine-grain details also. And we could
[29:19] also tell you which vertical is this happening because as Shailesh was showing the pipeline, right? Like we can look at each document, what are the segments that Databricks is being used and where Snowflake is being used. So, let's quickly check if this came back. Yes. So, this one shows you all
[29:37] the installs, right? Like we call that installs. Like if a company uses a particular product, we call that an install. And you can see that, right? Like I can ask it who are Databricks competitors?
[29:53] And what is the trend? Right? So, that kind of goes and it finds out it uses all our data based on installs. It will Yeah, it's just doing that, right? Like Okay, when it's doing that, I can show you into one of our interesting pipelines that Shailesh spoke about, right? Like for collecting data like
[30:09] this, we call that technographics. Anytime when you want to see which company is using what technology, that's generally the terminology is technographics. So, let's take this example, right? Let's say Northwind Logistics. I think
[30:25] even Google had this Northwind Logistics. So, they in their newsletter are saying that they are a proud Snowflake shop. But 6 weeks ago there is a job posting from them saying, "Hey, I'm a senior data engineer.
[30:40] Sorry, we are looking for a senior data engineer who could migrate us off from Snowflake to Databricks. And the resume says from a staff engineer in Northwind, "Hey, I migrated this." Right? Like so, there is a lot of breadcrumbs here and there. And we are this job of this
[30:55] Technographics pipeline is to make sense of that. As Shailesh spoke about it, right? Like we have like this whole concept of trying to do it cheap using uh keyword filters, uh BERT models, and uh like a increasingly progressive
[31:10] expensive uh step. So, I can quickly see if it did this. Okay, it's still going on. So, So, basically with the uh cheap filters, right? Like what we do is like uh try um uh use uh uh keywords so that we can extract a lot
[31:27] of useful information here. And then uh we use BERT kind of models to see uh how can we make sense of sentences, right? Some of these sentences are very useful. Some of them are not good. So, So, we try to reduce the noise. Because most of these web data uh has a lot of noise,
[31:43] right? Like 80% probably noise. So, then uh this is where our moat is almost there, right? Like all the 10 years of uh expertise that we have around GTM, we have distilled into these uh small Gemma models, open source models. And uh we
[31:59] try to extract not just the product, but context. And uh whether it is installed or not, right? Like those kind of subtleties are also extracted. And we are able to distill that here. So, if you look at it here, right? Like and we also look for use cases. Okay,
[32:15] where was Databricks used, right? Like uh interestingly, they are used in agricultural IoT sensor ingestion, security detection, retail. So, we could exactly know the company that is using it and also what kind of places that they are using it.
[32:32] And we don't stop with just Databricks using uh Databricks on names, right? Like we go into specific versions of it. If you could look If you see there here, like Palo Alto firewall networks, right? It's very uh specifically which uh router number or which version. So, this
[32:48] is all very critical for GTMs motions so that you know when to sell something to somebody when when they can be a security renewals and stuff like that. So, as I told you, right? Like, some of them say a document could say, "Hey, we There
[33:05] is no article experience required." That doesn't mean that they are using article, right? So, the SLM has to be smart enough to know that yes, this is not an install signal. This is more like a mention. And sometimes job descriptions will ask for everything in the
[33:20] ocean, right? They will say, "Hey, we need you to know about GCP, AWS, and Azure." But that doesn't mean that they are using all the three of them. So, these kind of subtleties are something that we have distilled into the model. Uh and there are things where the model
[33:38] may not be able to find out, and that is where we use frontier models, right? Like, we we use Gemini flash in this case. If you look at this statement, right? Like, it says, "Hey, we have a re-platform Cortex DMS onto Databricks." And there are few terminology here,
[33:54] right? Like, so the model is able to figure out that okay, this is some kind of transport management internal tool. And we use human in the loop very judiciously. We just don't throw everything at them. It's more like we do all the due diligence so that the next step the human can come in and say,
[34:10] "Hey, whether should I add this to our catalog? Is this something that people can know?" And the most important piece is like if something comes in all the way here, we make sure to train our keyword filters and SLM models so that this doesn't happen again, right? Like, human sees this kind of pattern only
[34:27] once. So, essentially, if you start from the beginning, right? Like, the company newsroom was saying something, job posting was saying something else, and the resume is showing something else. So, we kind of weight all of this and can come to conclusion like, "Okay, Databricks install is growing here. It's
[34:43] a Snowflake has been displaced in this particular case." So, this is like our entire technographics pipeline, right? And this is just one document, one company. We do this for like 7 billion documents and like 100K plus products, as Shailesh was saying, and 120 million companies,
[35:00] right? So, this is like our interesting technographics pipeline. Let me see. Yeah, it's back here, right? Yeah, so it says there are Snowflake seems to be the biggest one. Databricks, AWS Athena seems and with all the
[35:16] install counts. So, you could do a lot of analysis and you could ask for uh some specific questions around um what are the adjacent
[35:32] trends coming up and competition. Right, like you could keep playing with this, right? Like, so this one is um uh very good at that. But, okay, before before we go into that, right? Like, so
[35:48] the other most important thing is we are a very small company and this team is like very few folks and we cannot build all of the um platform and tooling for making this happen, right? Like, we use Google and Databricks heavily. So, I just wanted to quickly show what
[36:05] all we use on these two platforms. Like, for example, the Gemma open models, we fine-tune them using Databricks pipelines. We use the ML mod MLflow model registry and all our prompts are stored in MLflow prompt
[36:21] registry. So, we don't have to go anywhere else, right? Like, we use data asset bundles. right? is like within Databricks and we are able to have like the entire pipeline run in this. And just wanted to give a shout for the AI query because it does a lot of heavy
[36:37] lifting for us. As Abhishek already mentioned in the first slide, right? Like we do like we just have to call a sequel, right? Like we have like millions of rows and all we have to do is like call AI query and give that a sequel call and it just figure it it it takes care of
[36:54] batching, it takes care of checkpointing, it takes care of scaling all of the models. The LLM calls, right? If we have to do it on our own, it's a lot of API restrictions and rate rate limiting things that we have
[37:09] to do, right? So, all of that is taken care. Um so, yeah, we didn't have to build all of these orchestration layers and stuff like that. So, uh yeah, let's see. So, if if you want to use any of these stuff, we have agents and MCP and integrations. We can deliver
[37:25] the data to your S3 bucket and or your warehouse. And we also have some platforms like interesting. You can do sales co-pilot, you can do outreach. Uh So, these are all the ways you can consume Hidji's data. Let's see how this is going. Yeah.
[37:43] Yeah. So, yeah, there are like a lot of interesting emerging themes, right? Like for example, vector databases, new query engines, DuckDB. All of them are coming up with new AI-related products. Yeah. So, open table formats.
[38:01] So, yeah, that's pretty much it. We can keep playing with this. It It shows very insightful things and the way the use cases are being used. So, that's pretty much it. Thank you.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.