CDOIQ Symposium 2024 | Maciej Szpakowski | Prophecy

theCUBE.net

Clips
More from CDOIQ Symposium 2024

Maciej Szpakowski

Co-Founder

Prophecy

play_circle_outline Co-founder of Prophecy discussing challenges with data quality and data transformation

play_circle_outline Prophecy's unique approach with visual connected compiler interface

play_circle_outline Data Transformation: J&J and Texas Rangers Use Cases and Privacy Concerns with OpenAI Governance

Info
Transcript

Maciej Szpakowski | Prophecy

Maciej Szpakowski

Co-Founder Prophecy

The CUBE is covering the 18th annual CDOIQ event, formerly an MIT event, now independent. Dave Vellante and Sanjeev Mohan interview Maciek Szpakowski, co-founder of Prophecy, to discuss the company's mission and unique approach to data transformation. Prophecy offers a visual interface connected with a compiler, generating high-quality code from visual pipelines, with generative AI capabilities. The company plans to release a private model for specific use cases, addressing concerns of privacy and governance. Prophecy focuses on data transformation, solving d... Read more

explore Keep Exploring

What led to the decision to focus on solving the data problem before implementing AI solutions? add

What capabilities does Prophecy offer in terms of data transformation and pipeline building? add

What is the focus of the tool being discussed and how is it being used by Texas Rangers in their baseball games? add

bolt Powered by CUBE AI

Maciej Szpakowski | Prophecy

search

>> Hi, everybody. Welcome back to Cambridge, Massachusetts. You're watching theCUBE's coverage of the 18th annual CDOIQ event. It used to be an MIT event. Now it's on its own. This is theCUBE's eighth year covering this event. I'm Dave Vellante with my co-host, Sanjeev Mohan. Maciek Szpakowski is here. He's the co-founder of Prophecy. Good to see you. Thanks for coming on.

>> Yeah, absolutely. Great to be here. Thank you so much for having me.

>> You bet. So, co-founder. Why'd you start the company?

>> Yeah, absolutely. So there's the long story and the short story. I'll try to combine both of them together.

>> Yeah, give us the -

>> The middle version. Yeah, absolutely. So I think actually my previous company, we were working a lot on data, a lot on AI, building very specific models for pharmaceutical space, so very vertical-specific. And we kept seeing all the time is that even though you're trying to build a lot of fancy models, usually the bottleneck under the hood is the data. And everyone keeps talking about it for the last 20, 50, 100, however many years. Data is the problem, that there is a lot of problems with data quality. Very often, it comes down to the actual way you're ingesting the data and the data pipelines, the concept of essentially ETL. So in that company, we saw it firsthand. A lot of enterprises just struggle with their data. I think Raj Bains, who's our CEO, was one of theCUBE interviews as well. I've met him through a Berkeley SKYDECK accelerator for startups. We got to vibe a lot around, "Hey, how's difficulties to use for other people? It sucks in a lot of places." How data is just bad. And he was working at the time on a little bit more AI-facing company. We were like, "Hey, AI is great. If everyone's tracking with data, let's first solve the data problem." So yes, that's where we decided to join up the forces and work on the data space and data transformation.

>> The data is painful. It really is. Managing data, working with data, getting data quality is very, very hard to do, so I love the mission. But how do you do it in a way that's different and unique? What's your secret sauce? Is it math? What else is involved?

>> Yeah, absolutely. So we look at it as a combination of three different things. So first of all, Prophecy is a data transformation copilot. There used to be a lot of ETL, a lot of low-code tools in the last 20, 30 years. That's great. They've been working for, on-prem, they've been working on legacy systems. Most of the engineers hate them. They want to code up everything by themselves, but the business can't code. That's the usual story. And there's struggle. There is this conflict of IT is building one thing in code and then the business is trying to do another thing, maybe visually for Alteryx, Informatica, and all of those legacy tools. So the way we approach it is that our solution merges both of those worlds, and that's through our visual connected with the compiler interface. So whatever you do on the visual canvas, when you're building your data pipelines, they turn into high-quality code on the other side. Engineers love that. Engineers can modify that code and that then becomes, again, a visual pipeline on the other side for the users. Recently, over the last year, we also added generative AI capabilities into the product, so that actually a lot of the pipelines around 50% of the pipeline can be built just by the AI automatically for you to gain a lot of extra productivity.

>> So gen AI is bringing, what, the natural language capability, an orchestration?

>> What's an example of a pipeline that you can build conversationally?

>> Sure. So yeah, so there is two aspects to it. One is the conversational aspect to it. One is what we call promptless. Actually, promptless is what is really interesting for-

>> Promptless?

>> Promptless, yes, is what's really interesting to our customers, because with conversational interfaces, you're changing your mindset. You might be writing the requirements in English, but engineers, analysts are not really used to it. They live in the world of SQL and Python, so you're asking them to do something new yet again. With promptless interfaces, actually the tool just predicts the next step for you. So when you're typing in a column name that you expect, you get an expression on the other side for how that column should be defined. When you have a data set that you're trying to write to an existing table that you're trying to onboard new data sets on, you can actually just write a lot of those transformations automatically for that user too. So it's the combination of both of those interfaces.

>> So you call it promptless?

>> Yes.

>> But it's inferring a prompt from whatever-

>> Yeah.

>> And behind the scenes, are you doing multi-shot prompting but that the user doesn't see, or secret prompts that nobody sees?

>> Yeah, we would probably need a longer conversation to be back into all of that. So I think the part of the secret sauce behind it is definitely the knowledge graph that feeds in the details of that prompt, so we understand the schemas of the users, the transformation. And again, the tool is also a compiler. So we actually parse through all the code that the company has written historically, and we understand, hey, this is the common way that you're joining those two data sets. This is the common way that you're writing your expressions on the other side.

>> So are you training or fine-tuning an existing LLM?

>> Yeah. So at the moment, it's mostly OpenAI focused. We're using that as an LLM. It's the best LLM that is out there. And there's a lot of companies trying to build their LLMs, and then next day, OpenAI just does better. But we do have a story around building our own private model as well for very specific use cases. We see enterprises being concerned about privacy and governance in this space a lot. They don't want their data to be going to OpenAI, and it's natural, or their even metadata. So if they could be running a smaller private model that can be even more stable than OpenAI, you don't have to rely on an external endpoint, that's a very interesting story to them. We have something coming out later in the year for that.

>> The concern about OpenAI, just to double-click on that, is it's the quality of the results or is it the concerns over leakage proprietary data? Is it concerns over their lack of a governance structure? All of the above?

>> So Prophecy, we primarily target larger enterprises. That's where the most data is. That's where the biggest problems are. Fortune 500. Customers like J&J, for instance. Those guys do not want their metadata, their data transformations, or by no means their actual data, ending up in OpenAI hands. And it's not specifically about OpenAI. It's just about-

>> Any LLM.

>> Any LLM, exactly.

>> It's not OpenAI. But to your point, OpenAI, I would agree. From our own experience, OpenAI is the best. You've used others, but you get better results. Even personally, I use a lot of them, but OpenAI is pretty good. Are you building your own, you're saying, or building your product?

>> Yeah. So essentially, what we are seeing is OpenAI, you're right, is the best, but it's also for everything. It's not use case specific. With models, what happens is people train them on very specific things, and even with little data, you can make them outperform the generic models on those very specific tasks. So even though OpenAI is obviously throwing billions, those billions are on all the use cases under the earth or under the sun. So what we are trying to do is say, yes, we'll introduce a private model, but the private model is going to be trained for very specific use case like that next transformation prediction, expression prediction. So very small narrow use cases that you can optimize the model to be even better than OpenAI, or at least as good as OpenAI.

>> And that'll be your own LLM?

>> That'll be a fine-tuned version of an existing LLM. So we can just take the open source Llama 2 for instance, 7 billion parameter, and then optimize it, fine-tune it on a specific task.

>> So I'm going to go back to J&J example. So you were starting to tell us, so they don't want their data or even metadata to leave their premises, and you're using data transformation copilot that uses OpenAI's LLM. So how do you reconcile that?

>> Yeah, so essentially, at the moment, it's very metadata-focused.

>> I see.

>> And I want to just mention that the product itself, the copilot, it's not just about that AI piece. AI piece is important, but what a lot of people see these days is that AI gives you just code on the other side. And code is not suitable for vast majority of the customers. So even if J&J's business team would start consuming that code, they would just have no value out of it. You can ask OpenAI, it'll generate your code. It's garbage. They'll not be able to productionize it. They'll not be able to scale it. They'll not be able to even test it correctly. And that's where the other components of the product come in like the visual interface, where step by step you can actually understand what the pipeline does. And without having to understand how Python PySpark DSLs work, just define the data transformations by themselves.

>> So what are they doing with your product? When I look at what's being used in production, again, our survey partner, ETR, it's code generation, it's external customer support, it's text and data summarization, it's writing marketing copy. Most recently, we're getting collaboration, like summarizing meetings. Okay, we're getting more enterprise search is new, a little bit of image generation, and internal IT support and help desk, which maybe is from ServiceNow. What are they doing with Prophecy? Are these use cases or is it more specific?

>> Yeah, so those are I guess the general AI use cases?

>> Yeah, very chatty. Oh, yeah. These are the ones that say which business use cases have you put in production? And the ones I mentioned are the most prominent.

>> What business benefit is J&J getting from using-

>> Sure, sure, sure. Yeah, absolutely. So the tool, again, is primarily focused on data transformation, so it's not about any data summarizations or anything like that. It's coming back to the roots of essentially ETL and production data pipelines. It comes down to, hey, if you want to build an AI model in production or, like J&J is doing, reporting and analytics, you need really good high-quality data for it. So it's both the engineering teams and the analytical teams building and essentially data pipelines. Taking data from the sources, maybe it's clinical trial information in the J&J's case, and churning them into information that then can be structured and used for actual research.

>> Getting basically your data act together is what you are adding value to that process.

>> So it's that joining the data together, so that you can extract business insights from it.

>> So Dave, you're a big baseball fan.

>> Yeah, for sure.

>> One of their major clients is Texas Rangers, and they won the World Series. Yeah, they did an AI summit also, so can you tell us about Texas Rangers? How exactly are they using and what benefit are they getting out of-

>> Sure, sure. Yeah, so they're using Prophecy on top of Databricks. Essentially, when the game is happening or the practice of the game is happening, they're looking at all the data that shows up during it. They have sensors, cameras, iPads, recording all of that. And they're looking at how the pitchers and other properties of the game happened. When someone throws the ball, what's the angle of it? All of that is stored into their database sitting on Databricks. It's all data lake, but then this is raw data. So you need to now figure out, okay, well, you have all of those angles and you have scores maybe for those angles. When did you manage to hit the ball? Which are the ones that are actually going to work? So this is where they use now Prophecy to build that transformation on top of that data to find all of that information. Which are the angles, which are the properties, that are optimal for our players so they know how to improve their game and later on win the tournament?

>> This is a good use case.

>> Yeah.

>> I like it.

>> And they won the World series, so it's actually-

>> It worked.

>> Use Prophecy and you win the World Series. There it is.

>> It's the age-old debate about the nerds running baseball, but it's such a data-driven game.

>> Correct.

>> But I could see-

>> Oakland A's.

>> Right. Yeah, the Oakland A's the Moneyball. When you're talking about the angles, maybe it's like, okay, don't hit the ball if it's outside this zone.

>> Exactly.

>> Because otherwise, your probability of successes, or maybe, if you're defensively positioned-

>> Basically, whoever has the best data and the best analysis has an edge to win the tournament, right? It's not like you are the best players.

>> Well, this is what's interesting about the age-old debate, because the case with the Moneyball was it worked very well over a 160-game season. But when it came down to seven games, the better players seemed to always win, like in soccer this past weekend. It's like it just happened.

>> It is definitely not trying to replace everyone for this. This is all about that extra layer of information. But the people are definitely, it's all about the players for sure as well.

>> But it has a much, much bigger role in this specific example. So when did you start the company?

>> Yeah, so the company is around six, seven years old now. We're going to be getting into the seven-year-old mark. We have invested very heavily at the beginning at building the actual technology under the hood. It's not easy stuff. And then, again, we're working with enterprises. So one of our first customers was actually the biggest credit card provider in the world. So when you work with customers like that, they're pretty demanding in terms of the features, how you could deploy the product, et cetera. So we've spent a good amount of time just to see exactly what use cases they're building, what problems they're trying to solve, what challenges they're facing while building their data pipelines, and then optimize the product specifically on that. Over the last few years, we have released our copilot for Databricks specifically, so we had a big release also last Databricks AI summit, and we started giving it out to all the other enterprise customers.

>> So you've raised, is this right, 73, 74 million? Is that correct?

>> Yes. Recently raised series B.

>> And you feel like you have product-market fit at this point?

>> Oh yeah, absolutely. I think this challenge is repeatably seen across every single customer. I think very often what happens in the large enterprises is that they very heavily invested on the legacy side. They had those Initials, Informaticas, DataStages. Now they're trying to migrate a lot of that to the cloud, and they're seeing that the workforce is just not able to meet those new requirements. You need to know how the new world works. You need to be productive in it. You want to enable more people on it. And the process tool enables you to take all that code directly. One of our products is actually migration. A transpiler, we call it. So it takes all that legacy code, it spits you out new native spark code or SQL code that you can execute directly on top of your cloud data stack. And so that's the use case that every single enterprise these days is trying to move from legacy or make people productive on the cloud, so they see Prophecy as a huge benefit for them.

>> So modernization of ETL.

>> Exactly.

>> So you can take your old code, modernize it into spark.

>> That's right.

>> Run it directly against delta tables on Databricks.

>> That's right. Or even Alteryx, for instance.

>> Yes.

>> That's something that shows up very often. A lot of business analytics, how did it used to be done? Spreadsheets, Excel files, CSVs on SharePoint and SFTP. That just doesn't work anymore. If you're trying to build your AI model, you can't build it on CSV files, so now what you have to do is modernize that whole environment. And this is where we come in. We take that Alteryx mappings, we'll migrate them directly into your Databricks or Snowflake. You can start running it on your native cloud. And then you can keep on building more data transformations, whether as a business analyst, as a data scientist, machine learning engineer or data engineer, just directly on top of the Prophecy platform.

>> You called the product a transpiler?

>> Transpiler, yes. That's the technical-

>> A transformation compiler?

>> Yeah. We come from a little bit of a compiler background. The transpiler is essentially a source-to-source code converter.

>> Wasn't your co-founder working on Nvidia?

>> Yes. Yes, so that's Raj.

>> Raj, yeah.

>> Raj came from a very fun background. He was in Nvidia. I'm sure everyone knows Nvidia now because it's so huge.

>> What's Nvidia?

>> He was working on CUDA, which is the code optimizer that takes your C code and runs it on the GPUs. So he was one of the first engineers working on that, so that's where we get a lot of that compiler background from. But he was also in companies like Horton Works, prop PM for Hive, so very data-oriented. Microsoft as well, working on Visual Studio Code. So there is a lot of that background that comes in and plays around, plays together, in our company.

>> So your big competitor would be in Informatica, for example, or not necessarily?

>> I think we wouldn't really call them a competitor per se. They're in the landscape. They do ETL as well. But no one is comparing Prophecy to Informatica really.

>> You're not losing business to those guys.

>> Exactly. They're losing the business to us.

>> Ah, so they are competitive.

>> They're sitting on the legacy side. So if you're running Informatica on your on-prem systems and you're migrating them to cloud, you're usually not going to migrate to Informatica.

>> Informatica, by the way, just for the record, is trying to move to the cloud. IDMC is all cloud.

>> Sure.

>> They don't want to be on-prem anymore.

>> Sure, yeah. We'll see how they do that.

>> Yeah, no, no, absolutely. Is DBT one of your-

>> Yeah, so DBT is actually something that's used under the hood in the platform. So DBT is a build system for SQL, right? We are , for instance, for Java to build your code. That's the analogy for the engineers maybe out there. But then you also have DBT that builds your SQL code. So for our SQL product, we use DBT Core as the underlying build system for it. So that very often happens. There is a lot of code that has been written on DBT. People then struggle with maintenance of that code, so they can just bring in our product directly on top of that DBT code base and just allow you to keep on working on it, but in an easier-to-use visual way or code with those AI capabilities as well.

>> So Spark and SQL?

>> Spark and SQL. That's right, yes.

>> And Scala also, right?

>> Spark is both for Python where you have PySpark, and then Scala is actually where we started from. So our very first customer was working on Scala 4, but it's all Spark. It's all Spark.

>> Yeah, okay.

>> Well, congratulations, Maciek. Appreciate you coming on theCUBE. Great to have you.

>> Yeah, absolutely. It was great conversation. Thank you so much, guys. Thank you.

>> You're welcome.

>> Thank you.

>> All right, for Sanjeev Mohan, this is Dave Vellante. You're watching theCUBE's coverage of CDOIQ from Cambridge, Massachusetts. We'll be right back.