AI Pulse: June 2026

Santiago Giraldo: Hello, and welcome to Snowflake's AI Pulse, for the month of June. Just as a reminder, this is our monthly series where we showcase the latest and greatest innovations across Snowflake AI, Cortex, Snowflake CoCo, and Snowflake CoWork. Couple couple housekeeping things for you today. We're changing the format a little bit. Instead of doing q and a at the very end, we're gonna be doing a couple minutes of q and a after each section. This way, we keep things nice and tight. We can address the questions in a more timely manner when it's top of mind, and we can we can better better answer some of the complex questions that you might have around these specific topics. Regardless of that, during the presentation, please do not stop adding questions to the q and a. We'll continue to answer them either there or in the follow-up afterwards. Just a safe harbor and disclaimer statement, unusual stuff, take a moment to read. And then I wanted to quickly introduce our speakers for today. So my name is Santiago Giraldo. I am the senior director for product marketing, for AI and machine learning. We have Christine, who is our director of product management here at Snowflake Aria, our senior product manager at Snowflake, Qinyi Ding, our senior product manager at Snowflake, Arun Agarwal, who is our product marketing lead, for AI and ML. We have Avinash Joshi, our senior product manager, and then, of course, Effie, who'll be talking today about Snowflake CoWork. So, as we as we kinda dive in, I wanted to feature the agenda here for you. We're gonna start off with just a quick feature flash. I'm gonna talk a little bit about, what we're gonna be discussing today, as a whole. I'm gonna contextualize it in terms of the agentic enterprise for you. We're gonna take a look at availability, and then I'm gonna hand it off to Effie and Arun to dive into Snowflake CoWork, then Snowflake CoCo, adaptive compute, our latest innovations in agentic ML, optimized batch inference, and then, we'll have an additional q and a towards the end as well. So what's new? The coolest feature is hot off the press. So first and foremost, I wanted to level set a bit on how we actually view the world of the agentic enterprise today. What Snowflake is really setting out to do is enable the agentic enterprise for every organization, in the world. Basically, this constitutes of the agentic control plane that really, essentially captures the entire ecosystem of everything that's required to power AI driven workloads and drive AI driven, drive AI impact across your organization. This means having the availability to choose any model that's available, having, tools, for example, like Snowflake CoWork, which are really designed, for your knowledge workers to be more efficient and to get work done more effectively. Snowflake CoCo ultimately designed for that development life cycle and delivering results across the business more effectively. All of this wrapped together with your enterprise data, all of the context, semantics, and, and ontologies that are required to be effective here, as well as your ecosystem of software and applications. And wherever your data resides, being able to bring that all together and manage it in a way that makes sense for your organization in a central way that's governance secure. Today, we're essentially contributing to this vision by by basically expanding a lot of what we talked about at at Snowflake Summit earlier this month. Some of the things that we're gonna be talking about today are gonna be about the latest features in Snowflake CoWork, which is really now designed as your personal work agent designed for productivity for anyone in the organization. We're gonna be talking about Snowflake CoCo, which is really your development and insight partner, really designed for those builders inside of your inside of your organization. And then we're gonna be talking about new optimization features with things like adaptive compute, which essentially delivers self optimizing architecture and hardware for all of your organization needs, which ultimately lead to more efficient workloads, streamlined workloads, and cost savings. In addition to that, we're gonna learn a bit more about how all of these different pieces and the AgenTic enterprise contribute to delivering better predictive insights and more advanced, sophisticated, and novel AgenTic ML with things like, Cortex training. And then finally, we're gonna learn how all these pieces together, how we how the AgenTic enterprise as a whole, delivers production and scale, and can deliver basically an optimized, optimized batch inference and optimized delivery of outcomes and results for your organization. With that, just very quickly, I wanted to talk a little bit about our stat the status of some of these features that we're gonna be talking about. This is here just for reference for you. I won't dive into each individual item, all of these pieces are either coming GA soon or are gonna be available very shortly to everybody on Snowflake. And with that, I'd like to hand it off to Effie and Arun to take us through a deep dive of Snowflake CoWork. Effie, Arun?

Effie Goenawan: Yep. Hi, everyone. I'm really excited to give you an update on what's new with Snowflake CoWork. So here's a typical day of knowledge workers with dot a I. We used to spend so much time doing manial tasks, things like manually pulling raw data, drafting documents, decks, dashboard hopping. I used to have, like, twenty dashboard that I have to go through every single day. And then every time I have a a review meeting or preparing for a customer meeting, have to spend a lot of time building decks, building collaterals manually. And then, yep, after all these reviews get even more question that I have to spend even more time analyzing, filing tickets, waiting for other people to support me. And then end of day, I will spend the whole day. No decisions is made. It seems like no work gets done. So we are we at Snowflake, we believe the future of work is so much better. How we work must be agentic, proactive, and outcome driven, which is here's our vision of future of work. It's no longer human querying system. It's no longer me, you, and everyone spending a lot of time executing manial tasks. It's about using AI, using agents to supercharge us and our workforce, supercharging every knowledge worker so they can, elevate from doing this low level task to higher level, like steering strategy, oversight, and reviewing and approving actions rather than spending time in the weeds. So we are introducing this huge paradigm shift going from Snowflake Intelligence that we made generally available in November last year, which is talk to your Snowflake data, talk to your data intelligence to Snowflake CoWork. It's a shift for making Snowflake CoWork being your personal work agent, like your AI chief of staff. It makes CoWork be the starting point of every workday to get stuff done. It's not only reading and q and a reactive tool, but it's actually proactive. It's able to act. It's not just giving you information. It can automate the whole system end to end. It routes automatically. It can run multiple agenda tasks like deep research. It remembers who you are and learn what you like and your preferences. Your company is built on your company and context data. So then what's new with Snowflake CoWork? Tactically, three pillars. One is personal work agent. Personal agent, it's an agent that is yours, but, also, it's understanding your company and your business. So it knows the enterprise data. We at Snowflake, we we have the trust of your most sensitive data, your company's data, and then it routes automatically based on your preferences. It members and learns who you are, what you're working on, your company's context, semantic views, as well as business logic. And then it follows wherever you are, whether you are on mobile with iOS app, you're on Slack, or you are on desktop. And then second is my favorite, which is automation. So with CoWork, we have autonomous agent that does things proactively even before you wake up in the day, even when you are away from the keyboard. It automatically learns business context and produce deliverables of results, like finished deliverables. It's not only delivering you the answer that you need to copy place someplace else. It's also giving you a beautiful robust PDF, PPT documents that you can then bring to your meetings or your review. And then last but not least is our attempt for an AI first BI. We heard over and over that the dash people don't like or people don't enjoy swiping through multiple dashboards or spending a lot of efforts and time maintaining, building so many dashboards. Everybody kept on saying the dashboards are going away, but yet the next day, you go, you start the morning going to a dashboard. We believe that we are taking an AI first approach in an agenda era, building a governed dashboard end to end with natural language while keeping it governed and, shareable within an enterprise settings with CoWork. So I'm gonna walk you through each one of these three pillars. The first one is the personal agent. It's the one agent that connects all of the sub agents, the data agents that you might have built in the past, like sales agent, finance ops, products. Before Snowflake CoWork, you have to pick from an agent picker. Your user has to know which agent to talk to to get an answer. Now with Snowflake CoWork personal agent, all you need to do is just ask. We brought you to the correct sub agents, and then we bring in additional capabilities to get the work done, not just giving you an answer, but giving you PDF, PPT, anything collaterals that can be actionable as well as doing deep dive analysis to actually meaningfully move your business. We so we are focusing a lot more on the business outcome rather than just giving you insights. And then scales away for you to codify codify workflows. And then connecting with MCP tools, We know that work happens in multiple apps, multiple services, not just Snowflake. Work happens on email. Work happens on messaging apps like Slack. Also Jira tickets, Salesforce. We know we wanna meet you where you are, bring in all of the data that is most relevant to you to make you more effective. And then also on the personal agent side, we are bringing user centric capabilities like user memory. When I say, hey, I'm Effie. I'm product manager for Snowflake CoWork. Snowflake CoWork used to be known as Snowflake Intelligence. CoWork remembers who you are, your preferences, your role, and that helps you be more effective. So don't you don't have to repeat the previous context, including your preferences. Like, hey. For this kind of revenue charts, I like line charts or bar chart. Do not ever display pie chart. Coworker can remember that. And then user scale, it's codifying repeatable workflow. People have specific set of steps that they want to be done and they wanna distribute within their team or within their company. With Snowflake core, we allow you to codify this repeat workflow. I have customer meeting note prep or even, like, presentation prep skill that I can codify and run each time I have to repeat this workflow. And then I can share this useful skill to my team or my company so others can benefit in a covered way. And then last but not least is prebuilt skills, bundles of skills for top personas like finance and sales. So we help you get started from day zero to something usable within days, not weeks. And then on the personal side, we meet you where you are, your agent on the go. With iOS mobile app, you can use face ID to unlock. You can complete conversation history on mobile, pick up where you left off on mobile. This is really helpful for business users who are always on the go. For me, I'm always in between meetings. So I find mobile app to be really, really useful to get quick answer and also beautiful visuals. A lot of other AI tools have mobile apps, but we believe we have the mobile app that gives you the most robust and useful visuals on the go. And then automation, again, my favorite. I'm always trying to automate anything that I have to do repetitively. It saves a lot of my time. Essentially, set once, deliver always on schedule, meeting you in your inbox ready to act on. This morning, I wake up to an email telling me, hey, Effie. You have three customer meetings. You have AI polls. You have two meetings with your teams. Here is the questions you have to ask the customer. Here is a PDF I generated, and here are the top three things to prepare for the AI pulse presentation. Really helpful. I don't have to swipe through six different apps like I used to. I don't have to go to any dashboard. The insights come to me ready to act on. And then on automation, we also have sub agents capabilities with deep research and analytical search. The idea here is you have cohort agents spawning multiple sub agents to explore all of the possible ways, to answer your most complex wise, your most complex questions. For deep research questions like, why is my revenue dropping in September? What can I do to triple my business, my revenue? What happens is cohort can spin off multiple sub agents to explore all of the possible ways or all of the possible reasons why your revenue is dropping and what you can do or for all of the opportunities and come up with a one proposal for you to act on. So this is something that you would traditionally have to ask an analyst or, like, multiple analysts to explore for days, if not weeks, and pay a lot of money. Now you can do that in CoWork in less than fifteen minutes. So super easy, and I've heard customers love it so much that this is something where it's so unique to CoWork, helping you actually identify hidden insights that you might not even have thought of in the past. And then on the automation side, we have shifted CoWork from being an agent that produces things, not just answering questions, not a reactive tool. By that, I mean delivering PDF, PPT, advanced analysis, as well as customized actions. And, also, we add code execution so you can create skills that contain scripts for your users so they can automate that work a lot more effectively. And then on dashboard, our AI PA attempt, this is really popular among our customers. A lot of our customers are not satisfied with their traditional PI solutions. So they look up to us, hey, Snowflake, we already have our data with you. How can you help us visualize and facilitate collaboration? If you think about what is a dashboard, it's a means to an end. It's a means to collaboration, getting everyone on the same page, either on KPI, tracking people on the right track. So we are doubling down on artifacts. If you remember today, we have artifacts on CoWork. It's about to GA anytime now in the next few days, basically allowing you to get rich, beautiful visualization, whether that's a chart, any kind of chart, like hundreds kinds of chart that is worth saving and sharing. And the differentiator the way we do it is in a governed first way for enterprise. Meaning, when I shared the link with Arun, he opens it with his access. So Arun is not exposed to any data he's not supposed to see. So it's very clear. I can feel, safe, and admin can also have the peace of mind that people only see what they get to see. So I get this beautiful chart on cohort. I can quickly spin it on my home page. I can share it with Arun. He opens it. We are on the same page, but nothing is exposed that he's not supposed to see. Now we heard really, really good feedback on artifacts. You all wanna see multiple tile, more sophisticated artifacts, not just one chart. So we're bringing the dashboard, which is collection multi tile charts as an artifact. So analysts can create a North Star dashboard from Snowsight with CoCo natural language. I can say, build me my Snowflake CoWork consumption dashboard or my book of business dashboard. I can tweak it either manually with SQL or with natural language with CoCo, publish it to CoWork to a specific role, let's say salespeople, and then salespeople can go to CoWork and easily find the same dashboard with their privileges. And then if they have a question, they don't have to follow-up. They get for the analyst to answer. They can just ask directly. So this is, like, our first version of an AI first approach where everything is done through end to end, personalized to you, and, conversational. And then this is the artifact that I was referring to. It's a beautiful robust chart interactive that you can have a conversation with, save, and share. And then multi tile is a collection of these charts and visualization that analysts can create on workspaces published to cohort and have a natural conversation for the business users to follow-up with. With that, I'm gonna hand it over to Arun to give us a demo of how he's starting his day.

Arun Agarwal: Amazing. Thank you so much, Effie. Hi, everyone. So nice to meet you. My name is Arun Agarwal. I'm part of the AI product marketing team here at Snowflake. And for this demo, I'll be acting as a senior sales sales leader at Snowflake. And I'm really excited to show you how I start my day with CoWork and how I Snowflake CoWork is transforming the way I get my work done. So let's get started. You will see here that we'll let the video play. So here I start my day by getting an email from Snowflake CoWork where I can see some of the most important things about my customer. I can click on this link and I can directly get into the Snowflake CoWork interface. And here, I can start to get insights about my business. So you can see I can ask simple questions in plain English and I can start to get responses. I don't need to go ping my analyst or go to my data team. So here I'm looking at my top customer, Cornflower Kitchen, and I want to understand more details about their consumption, how the business is going. I can get that very easily. It can even create charts for me and visuals. Now this chart is interesting. I would like to look at it later so I can save it and pin it to my dashboard, and I can ask follow-up questions. Now I have a meeting with my customer tomorrow, and I want to really understand the business and make sure that I fully understand the different aspects of the business and how to grow it. So I can ask Snowflake CoWork for that type of preparation. Now because this is very comprehensive research, I can switch on my deep research mode, which is one of the features that Effie talked about. And what it does is starts to fire up my sub agents. It'll look through my structured data. It'll look through my unstructured data. It can look through a web search if it's turned on. And it's really this comprehensive type of research that would take weeks, if not days to get done. And I can do that very quickly using CoWork in a matter of minutes. In about ten minutes, we'll have this report ready. And what I can do is I can if I want to show you the output of it, let me get to the next slide here. Okay, so this is what I see after about ten minutes. It's the comprehensive report of Kwanflower Kitchen. My top customer, it'll show me which category has grown the highest in the past few months. It can show me how are my different products doing. It can show me different visuals that I can interact with. And this is really exciting. I think this is a report that I would like to convert it into a PDF so that I can either print it or I can preview it before going into my customer meeting. And what CoWork can do is it can kind of go through all of that information. It will package it for me in a PDF format. And, you know, it can also allow me to start to share this PDF. If I want to share it with Effie, for example, if she's coming to the meeting with me, I can use the same surface which is CoWork to kinda do all of that. So it's really exciting and powerful. I don't need to change my surface. I can stay in CoWork. I can get my insights. I can get deeper insights. I can download this PDF, and this is what the PDF looks like. Out of the box, super professional, very clean. It has all the information that I need. Obviously, if you wanna make it your brand color and all that, you can do that by customizing it. But really comprehensive report that I can start to use. And now I can ask CoWork to send it to FEI and email or Slack. And it can be done straight from Snowflake CoWork. And at this point, the email will be sent to the person I asked. Now the next thing is, you know, as a sales leader, I often care about my KPIs. I want to make sure that my business is performing and doing well. So for that, we talked about this. We have dashboards. These look amazing when you get to CoWork at the home page, you can find these things and now you can start to interact with them. So again, without reaching out to my analysts or my data teams, without writing a single line of SQL, without going through weeks of reporting. I can get insights into my data, get more curious to understand more aspects of my business through these dashboards, and I can use natural language to to query all of these things and get that information very quickly. So here you'll see that I'm asking some pretty complex questions about Conflower Kitchen and it's showing me all the details. It's showing me how the tools are growing, how the tool category is going, what is the total credit consumption for these different tools at different times. And all of that can be done very easily and quickly with a single prompt in Snowflake CoWork. So I don't have to leave the surface. I don't have to worry about wasting a lot of time away from my business. And now you can see it's showing me Conpower kitchen tool credit growth the last four weeks. I can click on it. Now what's interesting is I can further interact with it. So I can ask more questions on follow-up charts and those types of things. So really powerful tool overall. And then the last piece I wanna show you is, you know, as a sales leader, I'm not always tethered to my desk. There are many times when I'm I'm on the go. So for that, I can use the Snowflake CoWork mobile app with a simple frictionless login. I can use my face ID to log in and kinda get the same experience, the great experience I get on the web on my mobile. So this iOS app will be a perfect place for me if I'm, let's say, going to Conflower Kitchen meeting tomorrow. I can pull out my phone last minute, and I can ask questions like, hey. Let me understand more about their AIML usage. How are they using that? What's the latest consumption so that I can have a more meaningful conversation with my customer? And right there, can see all the information. You can easily get to insights. You can even share these with a team that was probably going with you for this meeting. So really simple and powerful, and this is truly the way Snowflake CoWork is transforming every knowledge worker's day and making it really more effective and efficient. So that's a short video video clips of our demo. With that, just to kind of wrap it up here for Snowflake CoWork, we have Snowflake CoWork in GA, so you can start to use it today. And you can use our quick starts to start to build more of your agents and then expose it into Snowflake CoWork. And with that, will turn it over to Santiago and Effie if you have any questions, we can answer it live. Thank you.

Santiago Giraldo: Okay. Excellent. Thank you, Arun. Thank you, Effie. And yes, we have quite a few questions. We're gonna take some minutes to answer answer some of these. And as I mentioned, we probably will not have time to get to all these, but we will answer these for you in either in real time or as we or in in the post follow-up here. So first and foremost, one of the questions that we have here is, let's see here. So one of the and this is an important question, I think. What makes CoWork better than something like ChatGPT, Gemini, or Claude? Arun, Efi, if you'd like to take that.

Arun Agarwal: Efi, do you wanna take that or do you

Effie Goenawan: Sounds good. Coming off mute. So CoWork so we offer you model interoperability. You don't wanna be locked into any specific model. So actually many, many ways to answer this question. Like, the first one that I hear resonate a lot with customers is the model interoperability. With some of these AI tools, you're locked into a specific model. And with Snowflake Core, we give you the flexibility and we are even delivering the model routing. So we automatically click the best model across different providers coming soon. So today, we know some models are better than the others, but things move so fast. So I a lot of our customers would want to have this flexibility, and it's very valuable. And then second is we offer the cloud based solution, so it's a lot. It's governed and secure rather than deploying a coding agent or like a desktop agent to business users, which carries a lot of risk and governance issues.

Santiago Giraldo: Yeah. I can I can expand on that a little bit more as well? In addition to being able to select, whatever agent is necessary or auto select that through, through dynamic routing, all of these models are running securely inside of your Snowflake infrastructure. This means that all of your security rules, governance, all of your semantics, your ontologies, and everything are inherited into Snowflake cohort, which means you can answer much more complex questions across your entire data estate as opposed to making any third party calls. In addition to that, because the infrastructure inside of Snowflake is designed to really give you access to all of that data, it lowers latency and makes it much more performant and all while maintaining that security and governance that's absolutely critical for running enterprise at scale. Okay. So this question, Effie, you mentioned inside of your presentation that Snowflake CoWork remembers and learns. And there's a couple of questions here about, how it actually does that. And if you could expand a little bit more around how that personalization actually works, I think it'd be valuable for this group.

Effie Goenawan: Yeah. We're adding capability for agents that have memory. And with CoWork personal work agent, we're rolling up user level memory. So we the the agent detect when the user is expressing their preferences or facts and then save it as the memory. And then we still also give visibility and configurability for the user to look at which memories are kept and then ability to delete and manage them.

Arun Agarwal: Yeah. It's almost like, you know, the demo that we showed. You can see for me as a sales leader, I will see Cornflower Kitchen in my dashboards, in my home page because that's one of my key accounts. If it is Santiago, you, it may be something else related to your campaigns that you're running as marketer. For Effie, as a product manager, it will be something about the products. So it's really personalized in that sense. You can start to see your most relevant information in CoWork and start to get insights by asking questions in natural language.

Santiago Giraldo: Excellent. Thank you. And I think we have time just for one more question as we before we dive into our Snowflake CoCo deep dive section. So one of the questions here is around around artifact creation. So the question here is when an agent oh, sorry. When when coworker is creating an artifact or creating a chart, is it the same across all the users? So if you ask the same prompt, will you receive the same answer? And then I think adjacent to that, from another question is, how does artifact sharing actually work, and what are some of the features that come along with that?

Effie Goenawan: Yeah. So let me break it down to a couple kinds of artifacts. The first one is a chart and a table, which is something we have in public preview right now. You can go to Snowflake CoWork, and you should see it in your account. That is allowing business user to share with another business user. So I can ask a question, show me my book of business. I get my book of business, and maybe it's a cornflour kitchen customer. I get this beautiful chart of cornflour consumption growth. I can save it and I can share it. And sharing it means creating a link, unlisted link that I can share with Arun. And Arun on his side, he clicks the link and that I share either through emails, like message. He taps it, and we render the artifacts running the SQL with his color swipe. So, essentially, I don't only sees what he's supposed to see. It's a live reference of the underlying data. And the second type of artifacts is the dashboards. This is like our major improvement to artifacts, which is not just one chart that the business user create, they can share. But it's actually a dashboard that the analyst creates on Snowsight with CoCo, publishing it to business users on cohort. So this is a complex, more complex visualization of the business insights. And when the user opens it, it's also respecting colors. Right? So Arun only sees what he's supposed to see, and, also, he sees the dashboard that is also shared with his role. So if the analyst shares a dashboard to a sales role and Arun is not a salesperson, he will not even see the dashboard. And even if he is a salesperson when he opens it, the the query what the data shown is only what he is allowed to see respecting RBAC and RLS.

Arun Agarwal: Right. And the data is real time. So if I if Effie sends me a chart, I open it. If I have access to that data, I will get the most refreshed information. I can start to ask questions and query them against natural language. It's really powerful. Think of it as I don't need to filter anything. I don't need to kind of go through a bunch of other dashboards. I can literally be in CoWork and get all of this information very quickly as long as I have role based access to the data that's underlying these artifacts.

Santiago Giraldo: Okay. Excellent. Thank you, Arun. Thank you, Effie. And that's the time we have for questions here. There's a lot of great questions. Hopefully, we can get to some at the end. But without further ado, let's dive into Snowflake CoCo.

Aria Attar: Hey, guys. Can everyone hear me? Beautiful. Well, nice to meet you all. My name is Aria Attar. I am one of the PMs on the CoCo platform. I formerly was cofounder CEO of a company called Tensorstacks, which was building autonomous agents for data engineering, and now as a part of the acquisition running a lot of the CoCo platform. Now over the last year, we've seen a massive adoption of coding agents across pretty much the whole software engineering life cycle. And one of the challenges within this adoption that we've seen is for data tasks and data engineering tasks. In a lot of cases, the tasks are a lot more rigid than software engineering, in the sense that if you ask Claude Code, for example, to go build out a front end component, there's a lot of correct answers in how you can build out a menu bar and have that menu bar work and have it be correct. And the feedback that we heard is in a lot of cases, that's not the case in data engineering because if you're aggregating together, you know, ten different tables and mapping a bunch of source data and so on, there's really only one right answer in a lot of cases on how you can aggregate it together to calculate some very specific business metric and so on. So what we needed was to go figure out a way to build a specific agent harness that's extremely good at this type of task with very high precision and accurate context. And that led to Cortex Code, which hopefully a lot of you guys have been enjoying up to this point. We'll talk a little bit more about how this harness works and how we've been able to achieve this in a second. But to kind of summarize a little bit over the last couple of months, CoCo has been a massive accelerator in terms of end to end development of the full data life cycle. Everything go from going from source data to modeling that source data, pushing it all into Snowflake through different pipeline tools such as Airflow and so on. And it's built around the same governance structure that you guys are used to with Snowflake. Now through this acceleration, we're we're also seeing massive adoption of CoCo for the end state of data as well. Right? So for everything on building machine learning models, building dashboards, and building end to end applications, Streamlit apps, and so on, which is really powerful. So you can take your raw data, model it, and then push it out to whatever that end state of that data looks like. Now, to talk a little bit about the COCA harness, and I'll explain what a harness is in a second. One of the most common questions that we get is, what's the difference between CoCo and something like Claude Code? And the answer to that lies in what the industry is now calling an agent harness. An agent harness is, you can think of it as everything around the model. So if you think about the base model such as GPT five point five or Claude Opus four point eight as like a car's engine, the harness is everything around that car engine. And this consists couple of pieces such as the system prompt, for example, that is a very detailed instruction set on how this agent should behave, how it should go look for data, how it should verify that some aggregation that it put together is correct, Down to the skills that it has, the tool space is a very important piece of the harness. The tools specify what capabilities this harness has such as, for example, Claude Code has the ability to edit files, read files, run bash commands and so on. CoCo can do all of that plus the ability to interact with Snowflake, and a lot of other internal tools that we built for for such as for like DBT specific use cases, such as the ability to trace a lineage of some DBT model basically instantly, or the ability to run semantic search on your giant snowflake schemas that you have, or the ability to interact with Airflow directly and so on. And this harness comes together to make it so CoCo's extremely accurate and extremely efficient at data tests. And this manifests itself in what we're looking at here, which is in every single eval set that we've basically ran for CoCo. We see something in the order of five percent to a fifteen percent boost in terms of precision to how accurate CoCo is at all these tasks compared to Claude Code and Codex, and something in the order of twenty to fifty percent less token spent per task, which is extremely important given how much costs are coming up in customer conversations. Now reflecting back on the last couple of months, CoCo ended up getting released back in November of twenty twenty five. And since then, we've been moving extremely fast to add new integrations into platforms such as DBT, Apache Airflow, new model releases, basically a day zero support for every new model that comes out. Exactly like Effie explained as well. CoCo is fully model agnostic, so you can swap between OpenAI Anthropic models, and very soon open models as well. And one thing that's very interesting and important is the harness that we build around GPT models versus cloud models are completely different, down to the way that they prefer to edit files. And this dynamically changes inside of CoCo as you flip between different models, which is really cool. So you you're not locked into one vendor, you can swap between any of them at any point. And more recently integrations into AWS Glue, Postgres, Spark, and very soon a lot more coming up. Now this has led to CoCo being, I think, this point, the fastest growing product that Snowflake has ever put out, which is extremely cool. At this point, we have around, honestly, over seventy one hundred customers across companies like Fanatics, Whoop, Morgan Stanley, and many, many more. I won't I won't spend too much time diving into all of that here. Now what we're gonna talk about today beyond this summary is all of the new innovations that we announced at Summit and all the cool functionalities that are coming up inside of Coke. All of these innovations can be bucketed into two core categories. Places that we're bringing CoCo to in terms of new surfaces and existing surfaces that you guys are already using today. And the second are programmatic ways that you can build on top of CoCo. Now, starting first with how we're bringing CoCo to the places that you work today. Up to this point, CoCo has been available across surfaces. You obviously have the CLI, which is a native binary that you can download, spin it up in any terminal across Windows, Linux, or Mac, and basically have it run fully on your machine the same way that a product like Claude Code would. You have CoCo in Snowsight, which has been up to this one a little bit more limited. People have been using it for a lot of like ad hoc queries, admin tasks, and so on. CoCo in VS Code is a new one that we announced fairly recently. So in whatever VS Code variant you prefer to use, whether it's cursor, native VS Code, you can just download the extension and spin it up. And lastly, the CoCo in plugin, which you can install inside of Claude Code or Codecs and have these agents refer to CoCo as a sub agent. So if you're already using Claude Code for your regular coding tasks, you can call you can have Claude Code call CoCo for any complex data tasks that it's doing, which is really cool. And we've seen massive adoption of this as well already, even inside of companies like OpenAI. OpenAI is using this CoCo plugin a lot, which is really awesome to see. Cool. Now, the next surface that we announced at Summit is CoCo Desktop. This is a new native application that you can install across Windows or your macOS machine. And this gives you the full CoCo experience just like CoCo CLI in a full desktop application with full visualization of charts, data pipelines, DBT models, and so on, which is really cool. It's very optimized for complex data engineering tasks, and this is also a VS Code. This is a VS Code variant as well. So all of the existing VS Code extensions that you use today, you can just directly import them inside of CoCo top and run with it the same way that you do. Now, CoCo Desktop has two modes. You have a regular IDE mode and a full agent mode. And the agent mode let's see if we can do we have a video? There we go. Yeah. We can watch over a demo of this. Perfect. Now agent mode is optimized for you to directly interact with the agent and offload a lot of the tasks that you would be doing directly to this thing without really worrying too much about what files it's editing and what the actual code base looks like. But with a click of a button, you can flip over to the IDE mode and get the full VS Code experience directly inside of the same agent mode, which is really cool. And inside of CoCo desktop, there's a lot of functionalities such as automations built in, which allows you to just like schedule repeated tasks if you want to, for example, load in data from some s three bucket into Snowflake. You can schedule automation where every morning it goes and checks for new data and goes into Snowflake, is very powerful. And by default, the same way that CoCo CLI is extendable by design, so is CoCo desktop. All of your favorite skills, MCPs, and ins that you're already using today are available inside of CoCo desktop as as well fully natively. Perfect. Now to move on to a couple of updates for CoCo and SnowSight let's go and pop this up. There we go. Previously, CoCo and we'll actually wait until this demo is finished, and then we'll jump over. Are we able to see the and SnowSight slide? Okay. Awesome. Perfect. So to summarize a little bit, in the past CoCo and Sosa has been heavily limited. If you guys are heavy workspace users, one of the things that you would have noticed is, as an example, CoCo CLI for some reason seem to be a lot more effective at very long running and complex tasks. While CoCo and Snowsight seemed to be a little bit more limited in terms of capabilities that it had. And the reason for this was CoCo and Snowsight did not have a computer to interact with. So it wasn't able to, for example, execute shell scripts or bash commands or run Python and so on, which is where a lot of the core capability that CoCo and c CLI kind of has, and that that's that's what makes us so powerful. On top of that, there was no ability to connect MCPs extended out with skills that required kind of arbitrary script execution like Python and so on. So in the past, CoCo and Sonosai has been heavily limited. And on top of that, if you're if you're like a heavy workspace user, one of the things you'll notice is you, for example, can't execute any bash. There's no terminal in workspaces. And that's part of what made CoCo and Snowsight so limited. Now with the announcement of Cloud Agents, that's completely flipped, which is really cool. Cloud Agents gives the ability to CoCo InstellSite to have all of the same tools that CoCo CLI did, which is extremely powerful. This gives a full cloud computer to CoCo and Snowsight, and CoCo and Snowsight is able to run Python. It could it can now execute JavaScript. It can literally do anything that it could inside of the CLI inside of Snowsight. And on top of that, what's very important is you can now have it run-in the background without your laptop being on or connected to the Internet, which I don't know if you guys have seen this, but at Snowflake in the Menlo Park office, you'll see people running around with a computer half open all the time just so their coding agent agent doesn't die off the second it they go in the elevator or they shut they shut off their laptop. This completely changes that. This allows these agents to run fully in the cloud in the background in more spaces without the need to have your laptop connected to the Internet, which is awesome. And on top of that, the way that Cloud Agents work is it's built on top of some new infrastructure we internally call CNG, which is a new really fast executing container service that allows CoCo sessions with a container to spin up sub second. And this infrastructure has been extremely transformative for us because it's allowed us to build a lot of very new cool surfaces, such as this CoCo mobile app and the CoCo Slack bot, which both are coming very soon. This allows you to kick off Cloud Agent sessions directly from Slack or a mobile app, and then resume it inside of Snowsight or resume it inside of the CLI or desktop, which is extremely cool. And this API, we'll talk a little bit more about that at the end of this as well. But that API will also be coming out externally very soon, so you guys can actually build your own Cloud Agents on the same API that we've put together. Last thing I'll cover here before we go into the programmatic ways that you can build on top of CoCo are gonna be skills and plugin sharing. So there's gonna be a new skills and plugins marketplace that's coming to Snowsight. This allows you to both internally and externally share your favorite skills that you've put together and plugins and fully extend all the different CoCo surfaces such as CoCo CLI desktop or now Cloud Agents in Snowsight. Cool. Now jumping over programmatic ways that you can build on top of CoCo. Up to this point, we've had two ways that you can do this. One has been the MCP server, and the second has been the ACP server. ACP has been a very useful way for people to build applications on top of CoCo. The use case for this is let's say you just wanna take the existing agent that CoCo is and just build a custom interface on top of it. You can use the ACP server to just spin this up and it works seamlessly. Interestingly enough, actually, own VS Code extension uses the exact same ACP server that's available to all of you guys. Now the CoCo SDK is the next version that's getting that's getting released. CoCo SDK gives you a programmatic interface to build agents end to end using a lot of the core components that we've already kind of put together such as assistant prompts, tools, and so on in Python and TypeScript. Now this SDK will also get the ability to spin up Cloud Agents very soon as well, which is really exciting. Now I'll pause here. I know we've covered a lot and we're running a little bit over time. So, Santi, if you'd like to kick off q and a, we can talk more about any of these surfaces.

Santiago Giraldo: Sure thing. Awesome presentation, Aria. We have time for one quick question because we are running behind time, and we have a lot a lot more great stuff to get to. But, Aria, I I think this is an important question that somebody asked. It's sorry. Just lost it. Okay. How do I know when I should use CoCo extension in VS Code versus something like CoCo desktop versus something like the CLI? For context,

Qinyi Ding: this user really good question.

Santiago Giraldo: Yeah. For context, this user today uses VS Code for all his coding, usually in the Cortex CLI.

Aria Attar: That's a really good question. What's beautiful about all these surfaces is all of it becomes pure user preference for you. Right? So whatever you prefer as like the surface that you're the most efficient with, our goal is to give you the same capabilities across all of them, and you get to just pick which one you prefer. I personally use CoCo CLI all the time for all of my things at this point. I don't have an IDE anymore. I just run-in a terminal. But co like, CoCo desktop has massive adoption, so does a VS Code extension. So it's purely up to you, which is awesome.

Santiago Giraldo: Awesome. Great. Whatever you're more comfortable. Thanks, Aria. Yeah. So there's a ton of questions coming in. We are going to answer these, but we're going to answer them either directly in the chat here, or as I mentioned, as a follow-up just so we can stay on track. And from here, we'll dive into adaptive compute. Thank you so much, Aria. Thank you.

Christine Avanessians: Thank you, everyone. Good morning. So you might be wondering why are we speaking about compute in an AI webinar today. Well, one will debate that the pace of AI innovation is unprecedented. However, it isn't just the AI models and the agents that need to move fast. The entire stack does, and especially the underlying compute infrastructure that powers it also has to keep up. Natural language inputs, as you're seeing when when you're working with with CoCo, dynamic SQL generation, what's actually happening behind the scenes when you're having these conversations, and just the diverse nature of the new data workloads are naturally spiky and unpredictable, which constrain these compute resources. For the teams to succeed with AI on enterprise data, they need a more dynamic approach approach to compute to handle the unprecedented scale and volume of these concurrent queries. So Snowflake's answer to this is adaptive compute. I am very excited to announce just two weeks after, our GA soon announcement of Summit when Christian Kleinerman was there on the keynote and we were talking about Adaptive, we have actually GA ed Adaptive on AWS. And we've also expanded the the region footprint, as you can see, to the six regions there. So we're going to dive into a little bit about what adaptive compute is and of your data workload needs. Adaptive compute focuses really on two major value propositions. One is on the ease of use. As you have many warehouses, how you configure them, how you manage them at scale becomes harder and harder. And we wanted to make a platform and a service that really made that even simpler than it is today. Simplicity has always been sort of a philosophy at Snowflake. And the second really is around high performance. We really wanted to push the boundaries and give you the absolute best performance we could provide commensurate to that cost. So let's really dig into the details. So what is adaptive compute? It's a new compute service. It's focused on this performance and effortless operations, as I was referring to. Adaptive compute still has the notion of warehouses today. These are called adaptive warehouses, a logical collection of queries. However, what is unique about adaptive computing and adaptive warehouses is all the jobs actually go to a pool of shared compute. So if you think about this diagram as sort of the environment of your account, the shared pool is within the context of your account. It doesn't span multiple accounts, doesn't span multiple organizations and other customers. And the goal is that this black box of the pool of shared compute really requires no infrastructure tuning. The system decides and ensures that you're getting the optimal performance you need workloads. So how much faster is adaptive compute? We like to regularly do our own experimentation and testing, you'll see some comparisons here, both to our standard generation one warehouse and our standard two generation warehouse. We are up to one point six times faster for analytics, two point two times more throughput for AI and BI operational analytics, and three and a half times faster for DML heavy workloads. These were measured using a combination of industry standards like TPCDS and our homegrown benchmarks. And so you can see we started off with our generation one. We made a huge push last year with generation two, but we really kind of wanted to reinvent and push ourselves harder, both on the performance angle, but then also bring in that element of abusive use. So many people ask me, so is there anything I'm supposed to do as an end user? We do provide two specific parameters. A lot of the things that you do today when you think about sizing your warehouse, scaling up, scaling down, using multi cluster warehousing capability, query acceleration services, when to auto suspend and resume. All of these details, they're taken care of for you by the system. The scale up, the scale down, the scale out, the scale in, all of those we will ensure to figure out what is the optimal compute, quote unquote, for a particular query that comes in. But we understand that usually users want to provide some input. You know your workloads best ultimately, and you can think of these as guidelines or hints that you're providing to the system. So the first one really is around your performance level. This is an expression of the upper bound of performance for any individual query. The way to think about it is really you're telling us how quickly should we really run. What is that performance profile you want for the queries that are coming up? It's a little bit of a speed up, speed down relative dial, so to speak. It's expressed in t shirt sizes. However, I know many people might think of the t shirt size and automatically think about a particular compute cluster. It's not quite like that. The t shirt size conveys a similar or better performance level than its commensurate classic size. So in the example, if I set max query performance to large, and I have a couple of jobs coming in, let's say job one the system decides really could benefit from performance. We would still respect the setting that the user has set here at the warehouse level to large, and we'd run it on large. If job two comes in and it's small and we determine that it needs a small performance level, this is an upper bound. So we wouldn't automatically default it to large. And that's sort of the starting point of this notion of the system really figuring out what your query needs and trying to create the back end compute for that. Speed or how quickly something goes through your system, usually there's the second part to that, which is how much that goes through the system. And so we have a second property, the query throughput multiplier. This really focuses on it's a scale factor, partly because there's a system baseline we didn't want users to think about. But it helps control how many queries are the total work on the adaptive warehouse that can be run at any point in time. We take these two parameters. We come up with a budget. Every query comes in. We assess how much compute it needs. We take away from the budget. And when there's no more budget, things start to queue. And once the system drains, then we take on more queries. So if you set it too low, the system will sort of it will be a pressure, and you'll start seeing queuing. If you set it high or if you set it to infinite, which is zero, you will the throughput really will be the throughput of your workload. It'll be defined by that. So in the diagram, you can see when I set it here, I started seeing queuing. When I set it to a larger value, then it's really the blue line, which is my workload. So I know we're kind of going quickly a little bit to catch up on time. I will have a demo that kind of walks us through this. Creating adaptive compute is super simple. You have two ways of doing it. You can just create a net new adaptive warehouse, and you can set the properties directly, or you can allow the system to to default. You can also do an ultra warehouse. If you have a gen one or gen two warehouse, you can easily, convert it over to an adaptive warehouse. No downtime. That's genuinely true internally. When when we, do our testing, we switch back between adaptive and classic warehouses on a daily basis. So we are in production doing this back and forth for our own internal data warehouses as well. If you don't set the values of the parameters, the system will evaluate the settings of the classic warehouse. So I'm going to jump, ahead. Very quickly, the adaptive warehouses use a query based billing model. What that really means is we bill on the compute and software used by each query. I'm going to jump ahead here. We are also announcing that we will have an experimentation framework similar to kind of the methodology we use in public preview soon. It helps you create these experiments, monitor, and analyze them. I want to jump into the demo here. I'm not seeing it playing. Anyone can help me as to why this is not laying? Oh, perfect. So in this demo, what you can see is we're creating a new adaptive warehouse from Snowsight. In this example, we are changing the parameters that we talked about. You can see that it's creating the warehouse. On this front, it looks pretty similar. It is still a warehouse like you would today. We are going into looking at the warehouse itself. There are the properties at the very end when you list the warehouse. We are also, going to use the adaptive warehouse here, and then we are actually gonna do a a quick usage of CoCo. It's asking about some properties of the adaptive warehouse, and it will also change those properties via CoCo. So if you don't want to use the SQL statements directly, you can very well use CoCo. I was doing a similar presentation about a month ago. I needed to create a huge demo database. I was using CoCo, and it was helping me as I was creating my my demo as well. And as it's as it's finishing up, I'll briefly describe what we're gonna do next. When you of the next set of code, we had a couple of warehouses that were running. One day, it was running classic. One day, it was running adaptive, and we were switching back and forth. And I we were trying to compare the p fifty performance. So this piece of code is gonna go query the right data sources in in the account usage views. And then it's going to demonstrate a chart as we're kind of showing this. It's also going to highlight the differences. So you can see here the orange bars were a classic warehouse. The blue bars were adaptive. You can see speed of web. Then finally, what this shows is there is a new usage view that is available that will give you the per query insight. So this is just querying it and showcasing those details. We do have blog posts, documentation, longer videos, and all of that. So with that, we can actually open up to Q and A if there are any questions.

Santiago Giraldo: Awesome. Excellent. Thank you, Christine. So I think we have time for one question as we have two more very exciting topics to get to. But one of the recurring questions here is around cost. How does cost for adaptive compute virtual warehouse compare to that of a standard Gen one?

Christine Avanessians: Yes. So it is a different billing model, as I mentioned. But our goal is to actually give you better performance or more queries than Gen two at a similar cost to Gen two. This is also why we've been doing that extensive testing. Of course, mileage varies per workload, but we are trying to give you more performance for similar cost in a slightly different billing model.

Santiago Giraldo: Awesome. Thank you, Christine. Great presentation. So now we're going to dive right into Snowflake ML, another great and exciting topic coming from our our our announcements this summer. So take it away.

Avinash Joshi: Oh, thanks, Santi. Hello, everyone. Excited to be here representing the Snowflake ML team. I'm a senior product manager here at at Snowflake on the ML platform. I'm here to talk about agentic ML. Now this slide we had seen already in the earlier in this webinar, but it's more so relevant for for ML, where you are in the previous world, at least for me, I I started off as a data scientist. And my my initial years, I remember more than sixty percent of the time, I I'm spending my time data wrangling, preparing my environment, dealing with Python dependencies, etcetera, instead of actually building the model or solving the business use case. Now with and now with the agentic ML framework and using CoCo, the the script is flipped. Now most of the times, my time is spent in steering and working on the strategy that I want to deploy to solve my business problem by by designing experiments, running through iterations, using my domain expertise, and and all I'm doing now is doing workflow oversight where the grant work is being done by by the agents. How do we do that? Is meet CoCo. CoCo is your ML agent with full context of your tables, your schemas, your databases, and your models that have already been built within Snowflake. When you use CoCo, it's not just another agent. It's different from Cursor or your Copilot of the world because it lives and bleeds within your data platform. It knows your schemas. It knows your tables. It knows your notebooks, and it knows your business team's best practices. So when you say, building a churn model using this my transactions table, it doesn't start with first, give me a give me the date table name, give me the database. It just works. It doesn't just build this task a scaffolding code for you, but actually builds a notebook, runs it in a secure sandbox, iterates through it, and make sure make sure that it works, validates the results, and then gives you an end to end solution for you for you to start then iterating over. Using this using Snowflake CoCo, you can use an entire ML platform, which which Snowflake ML platform, which is a natively context aware platform across data, models, notebooks, features in a unifying governed environment. This is Snowflake ML. You can use Snowflake ML through either CoCo or directly through notebooks or external IDs. Using these interfaces, you can access the the entire feature set of audio models life cycle, starting from development and iterating where you have your Snowflake ML APIs, which allow you to train your models in a distributed fashion, Experiment tracking, which lets you log all your experiments as iterations and runs, which you can use for then comparing your comparing across your different runs and finding your best versions for your models. And cortex training, which helps you train, fine tune, and post train your LLMs. Next is orchestration automation. You can use ML jobs functionality for setting up your training pipelines in which going through from both preprocessing, training your models, and post processing to orchestrate the end to end ML pipeline, which results in a model. Once you have your model, you want to manage that through model registry. So your model registry is a single pane of glass where all your models that you want to deploy to production end up. Using model registry, you can manage the life cycle of your model where you've where you've gotten the model from a development environment, you run your tests, and then get it staged for a production deployment. You can also manage your features features using feature store, making them ready for your by man by making by promoting reuse and then making sure that the features that are being developed are collaborative and used across your teams. The models in model registry and features in Feature Store can then be deployed and served in production in multiple factions, such as in batch mode for your daily workloads or real time for your real time needs. And finally, anything that is deployed in production needs to be measured and monitored. For that, we have ML observability and explainability functionality that allows you to make sure that the models that are deployed in production are performing expert as per your expectations. All of these can be deployed to either to contain a runtime or to ML ML optimized warehouses. Let's talk a little bit about each of these training feature store and inference benchmarks. So in a first, quickly, what what training what training lets you do is with these tools, training is a lot more flexible within Snowflake ML. You did you know that training is a lot more performant and but did you know that training is a lot more performant and cost effective too with optimized data injections, training pipeline optimizations, and uniform memory architecture? Training models like PyTorch and Snowflake is up to two point five x faster and three x cheaper than our competitive platforms. And the key addition of CoCo is makes it makes using this platform very easy. You just tell your objective to a CoCo, and it picks the training algorithm and tunes it for you. You define the goal. It finds the right approach. Real time features and model serving is the next new fee the it's the newest feature that we launched in the, in the last in the last month where you have your low latency predictions, which serve features under thirty milliseconds milliseconds in and models that serve predictions under hundred milliseconds to power online use cases. We have natively integrated MLOps, which will let you easily monitor and explain and receive alerts when the model performance degrades. All of this happens within the unified security and governance boundary of Snowflake for both your online and batch inference suites. Again, I like training for info feature store when it comes to serving a real time features features. We beat up we beat competition hands down where our q our when for parallel queries per second up to two hundred queries per second, you can see that we are much lower in latency than our competition. And above all, as as in when these number of queries per second go go up, we don't fall off. Suppose this backed online feature store that powers this this performance with ten millisecond. At p fifty, you can get under ten milliseconds of feature serving. Finally, inference, which is where everything comes together, where we have the best in class price to performance ratio for for inference, where we can run up to five x faster inference and up to three x cheaper by eliminating data movement and maximizing your compute throughput. You can deploy low latency inference endpoints with just a few clicks and easily deploy models for high throughput inference. And we we configure the state of the art inference engine based on whether it's a traditional model or a VLLM or an LLM model. You can also deploy models at batch where we we can scale up to billions of rows daily of daily predictions with optimized batch workloads, which you will speak which we'll be talking about in the next segment. And, again, as you must have heard the theme of this entire session so far, CoCo handles all the deployment and the configurations with just a single instruction. And with that, I'll give it back to Santi. Thank you. I'm happy to take any questions.

Santiago Giraldo: Excellent. Thank you so much. I mean, great presentation. I think the Jintiq ML capabilities are really quite impressive, and it's really the future of a lot of the work that we're doing around scaling machine learning workloads and solving novel problems. So let's take one question here. And this is good question. What is the accuracy of Snowflake ML models for things like forecasting? And I I know that that can be largely arbitrary and dependent on data, but maybe we can talk a little bit about how we handle accuracy and how how a Jentgen ML plays a role in that.

Avinash Joshi: Happy to. Thank you. That's a great question. So we think of us as a Snowflake ML platform where a data scientist would come and build their own models. So so the accuracy of the model is largely dependent on what sort of quality of the data, what sort of practices they're performing. But the Genetic ML, what it lets you do is it lets you rip through multiple experiments and iterate very quickly. So if you need to build something, some say a forecasting model, and you want to try different time lags in your variables, all you have to tell CoCo is to generate ten different training datasets with different time lags. It'll it'll do that within seconds for you, and then ask in parallel, start a training job in which and then how you have ten model study, which you can then start comparing using our experiment tracking feature. So that's how I would think about accuracy here, where you CoCo our platform, along with CocoM, lets you experiment very quickly and rapidly, where before doing something like this would probably take you like a week. Now you can do that within an hour. Excellent.

Santiago Giraldo: So fantastic. Just moving right along, looking at time. Let's go ahead and jump over to the next section. And we're going to be talking about optimized batch inference. You so much, Avinash Joshi, and, and we'll dive right in.

Qinyi Ding: Thank you, Avinash Joshi. Thank you, Santi. Hi, everyone. My name is Qinyi Ding. I'm a product manager of the Snowpark team. In this section, I'd like to give a deep dive into the optimized batching first workload. So first, I'd like to give an overview of Snowpark. Snowpark is already used at scale by customers building Python, Java, and Scala pipelines directly on Snowflake. Today, it powers three hundred fifteen million plus daily jobs with zero infrastructure for customers to manage. We see customers using Snowpark across a few common patterns, data pipelines, and our batch inference, programmatic orchestration, and unstructured data processing. I think the big idea is simple. Bring your code to your data and let Snowpark handle the scale, governance, and operations. We have been working to make Snowpark even more amazing. I'd like to highlight three key areas with new features. On developer productivity, we support proxying your packages from your own private repositories, provide comprehensive CoCo skills for Python script authoring, deployment, and observability, and hybrid pandas to automatically switch between pandas on Snowflake and the vanilla pandas. On Python data pipeline and orchestration, we support ability to connect with external data sources through DB API, JDBC, and Egress IP. Our ML and unstructured workloads, we support UDF to stage rights, directory import, optimized batch inference on UDF, and AI SQL functions. So together, these unlock more end to end Python workflows directly inside Snowflake. So for example, we see customers using Snowpark for ML batch inference at scale. A typical workflow looks like this. Ingest structured data and unstructured files into Snowflake using Python packages and ML models of their choice by artifact repository and directory imports, feature engineering with Snowflake ML, Snowpark data frame, and UDF, and then run batch inferencing at Snowflake scale and reliability using model registry. This is useful for things like document classification, customer recommendation, fraud detection, and other large scale prediction workloads. So next, I'd like to show you a Snowpark Snowpark based fraud detection pipeline to showcase the capabilities.

Santiago Giraldo: Gonna share my screen.

Qinyi Ding: Showing up my notebook. So here's a scenario. We are a financial services company. Our transaction processing system runs on Postgres. We need to get transactions out of Postgres, engineer features, and score every transaction for fraud, all inside Snowflake using Python. The four features I like to demonstrate are DB API for ingestion, vectorized UDFs for feature engineering, model registry for optimized batch inference, and customer hosted artifact repository for package management. So here's how the pipeline architecture looks like. In step one, we use DB API to get our data out of Postgres with a standard cycle p g two connect connection factory. In step two, we use vectorized UDF to engineer features, run it in batches instead of roll by roll. In step three, we use model registry to make batch inference simple and efficient. We will register a locally trained XGBoost model, and then it will apply optimizations for all UDF workloads to share the same model. In step four, we're gonna apply the model to the enriched dataset that we have prepared before and give each transaction a fraudulent score between zero to one. Throughout the demo, all the Python packages are pulled from our in house Nexus repository using the artifact repository feature. So before we go into the pipeline, there is a one time admin setup. We create an external access integration so Snowflake can reach out to Postgres. This involves a secret, a network rule, and external access integration. There's a similar pattern for Nexus integration. We need a secret, API integration, and artifact repository object. All these are a one time thing that account admin should configure. So I have preconfigured all these objects, so I'm gonna skip this step. So the first step comes from our transaction data living in Postgres. So traditionally let me run start running this out first. Traditionally, that means setting up third party ETL tools to move data into Snowflake. With Snowpark DB API, you write a standard cycle PG two connect connect call and pass it to the session dot read dot DB API, and Snowpark handles the rest. And we wrap this whole thing on into a stored procedure by the add spark annotation so the DB API connects from Snowflake directly to Postgres to fetch data. So a few things to cut out. You can get data from a table or a query. You can specify the number of partitions, in this case four, which means Snowpark will split the reads into four, parallel workers, each pulling a range of transaction IDs. You can also specify the fetch size, which which means how many rows to write in each file. As we can see, fifty thousand transactions are are being ingested from Postgres. And looking at the data, it's a typical transaction data with customer IDs, merchant IDs, amount, and time stamp. Now we're going to engineer the features. So this is where the vectorized UDF comes in. Traditional UDF calls your function once per row. Fifty thousand rows means fifty thousand Python calls. Vectorized UDF flips this. So here, by looking by specifying the panda series of input and output, SoapRoc knows to register this as a vectorized UDF. This function receives batches instead of individual values. In this case, we will calculate the the z scores, cyclical timing timing encoding, and category risk lookup. Notice that, you know, in artifact repository, we're specifying my Nexus repo, which we configured before. So all these packages like pandas, NumPy, topical are coming from our Nexus repository. After we register, the UDF, we're gonna apply the UDF on top of the transaction, table that we pull from Postgres and, provide enriched feature, feature columns. And also we're joining our custom profiles. So that's a, feature table created again with fifty thousand rows. So now we're going to create, the scoring models. By the way, this is what the table looks like. We have the existing columns with z scores and our encodings and, category risk in, and a bunch of other flags. Going into the next step, we're gonna score the model. We train a XGBoost model here, a notebook, and we're gonna going to register it, using model registry. Gonna run it now to let the model start registering. Notice that there's options here, with model in it once. This tells the model registry to load the model instead of into the head worker so that all the UDF workers share the same model instead of loading them locally. As the model finishes loading, I'm gonna then apply the model onto the feature trained dataset that we have previously prepared. At this step, we are scoring all fifty transactions. We compute the transaction velocity as in how many how many transactions per customer in the last twenty four hours, and then we're gonna apply the model run, to run the model that we just registered. And as a result, every transaction has a fraud probability between zero to one assigned to it. So lastly, as a final step, let's see what the model found. So these are the top twenty most suspicious transactions. These are high amounts, repeating customers showing up in multiple times, low risk customers suddenly making big purchases, stuff like that. In this particular example, you can see customer four ninety four showing up three times with one transaction of four thousand three hundred dollars in the grocery, which, you know, looks very suspicious. Again, that's something that the fraud analyst would investigate further. So just to recap, in this demo, we used Snowpark to build an end to end Python pipeline from external connectivity to transformation to ML scoring with enterprise package governance built in. So with that, that concludes my presentation, and thank you very much.

Santiago Giraldo: Thank you so much. I think we have time for one quick question. I I I think that this is an important one. When we're defining optimal, and, we're we're defining optimized, you mentioned that this is really defined in terms of performance. Are there scenarios where you can optimize for cost instead or make sure that, that costs are are regulated in in a certain way as you're working with these systems?

Qinyi Ding: I, I'm not sure if Avinash Joshi is, still on, but, you know, I think there are, features in the model registry that is, doing that. But, you know, I would like to defer to him for, the cost optimization part.

Santiago Giraldo: Okay. Perfect. Second here.

Avinash Joshi: Yeah. I'm I'm here. Sorry. What is the question? And if unless unless you answered it already, then we can move on.

Santiago Giraldo: No worries. It it was around optimization. What what features do we have around cost optimization, especially as we Yeah. Talk about optimization.

Qinyi Ding: Yeah. I mean, Gokul can

Avinash Joshi: give you an estimate of what it'll cost. Like, it can tell you how long each job is gonna run. So, ultimately, Snowflake is a pay ML platform is primarily a pay for pay for compute model. So the faster your job ends, the lower the lower the cost. So you can always that's and you can ask Okur to find me the model that will not only be performing the best, but optimize between performance and cost by finding me the model that'll that can run, say, ten million predictions with what's what's which model will run ten million predictions the fastest. Go ahead give you an estimate for it.

Santiago Giraldo: Excellent. So here's another question. Is there any tool for scheduling the jobs within SF ecosystems?

Qinyi Ding: I think, typically, we are going to tasks and task graphs, for scheduling jobs, and I think that's perfect for, like, securing, you know, slow procedures and anything like that.

Avinash Joshi: And one other thing to add there would be if you have a single table in which your everyday data comes in, like it's an incremental data, then dynamic tables are also amazing at doing, this kind of work where you have you want only predictions to have say, for example, you have ten million rows coming every day. So so you have day one, ten million rows have come in. You run your model batch inference. And the next day, another ten million come in, and you want predictions to only run on the new ten million rows. Dynamic tables are exactly designed for that. That's another option.

Qinyi Ding: Yeah. That's a great point. On the declarative pipelines that, you know, dynamic table is perfect. Excellent.

Santiago Giraldo: And then, last question here. Folks are really interested in the demo that you gave. Is there a GitHub repository or a case study or downloadable notebook that we can share on that?

Qinyi Ding: Not yet, but, you know, I'm looking into, like, publishing that, soon after.

Santiago Giraldo: Sure. So we can make sure that I send that out as a follow-up. Okay. So with that, this wraps up our, our session today, for Snowflake's AI Pulse for June. Just a reminder, this is a month, so be sure to tune in, next month where we're gonna be diving more deeply into things like AI functions, Snowflake CoCo, and lots of other innovations that are coming really fast out of the Snowflake team. Thank you all very much.