UiPath Fusion 2025 | Jerry Liu, LlamaIndex & Taqi Jaffri, UiPath

Clips
More from UiPath Fusion 2025

Jerry Liu

CEO

LlamaIndex

Taqi Jaffri

Sr. Director, Product Management

UiPath

UiPath and LlamaIndex team up to bring orchestration to the era of agentic AI

As artificial intelligence agents become a part of the business workflow, the orchestration of all the different human and nonhuman components grows more complex.Business automation platform company UiPath Inc. aims to simplify the process of building autonomous agents for enterprise. It supports multiple frameworks, including LlamaIndex Inc., which offers a toolkit for developing AI agents based on contextual data.UiPath’s Taqi Jaffri and LlamaIndex’s Jerry Liu discuss agentic AI with theCUBE.“People think orchestration is just stitching together different systems and calling it a day,” said Taqi Jaffri (pictured, left), senior director of product management at UiPath. “It’s not

play_circle_outline Introduction of guests Jerry Liu (LlamaIndex CEO) and Taqi Jaffri (UiPath Senior Director).

play_circle_outline Overview of LlamaIndex's mission to connect enterprise data with AI technologies.

play_circle_outline UiPath's open-source strategy to foster innovation and build versatile agentic systems.

play_circle_outline The challenge in maintaining stable enterprise-level open-source solutions amidst rapid AI advancements.

play_circle_outline Enhancing Automation Accuracy: The Collaborative Potential of LlamaIndex and UiPath in Building Complex Enterprise Workflows

Info
Transcript

Jerry Liu, LlamaIndex & Taqi Jaffri, UiPath

Jerry Liu

CEO LlamaIndex

Taqi Jaffri

Sr. Director, Product Management UiPath

At UiPath Fusion 2025, theCUBE’s Rebecca Knight sits down with Jerry Liu, CEO & co-founder of LlamaIndex, and Taqi Jaffri, senior director of product management at UiPath, to unpack how enterprises are moving from legacy RPA to intelligent, agentic automation. Liu explains how LlamaIndex evolved from solving early context-window limits into a developer and managed services toolkit for connecting massive, unstructured enterprise data to state-of-the-art models – powering multi-step, human-in-the-loop workflows far beyond simple RAG chatbots. Jaffri outlines Ui... Read more

explore Keep Exploring

What is the background and current role of Jerry Liu in the context of UiPath Fusion 2025? add

What is LlamaIndex and how was it developed? add

What is UiPath's open source strategy and how does it relate to LlamaIndex? add

What has surprised enterprises about the usage of LlamaIndex? add

What are the goals and expectations for the LlamaIndex and UiPath partnership in terms of document workflow automation? add

bolt Powered by CUBE AI

Jerry Liu, LlamaIndex & Taqi Jaffri, UiPath

search

>> (Music)

Rebecca Knight

>> Good afternoon, everyone, and welcome back to theCUBE's live coverage of UiPath Fusion 2025. I'm your host, Rebecca Knight, alongside Dave Vellante, my co-host and analyst. We've got two great guests for this next segment. I'd like to welcome Jerry Liu. He is the CEO and co-founder of LlamaIndex. Welcome, Jerry.

Jerry Liu

>> Thanks for having me.

Rebecca Knight

>> And Taqi Jaffri, senior director of product management at UiPath. Welcome back, Taqi.

Taqi Jaffri

>> Thank you. Nice to be back.

Rebecca Knight

>> You were here last year. Great time, yes.

Taqi Jaffri

>> Thank you.

Rebecca Knight

>> So Jerry, I'm going to start with you. You are the co-founder of LlamaIndex. It's a two-year-old company. Tell us a little bit about LlamaIndex to set the stage.

Jerry Liu

>> Yeah, sounds great. So two and a half years ago, or almost three years ago at this point in time, GPT-3 was out and a lot of developers were tinkering around with LLMs to try to make sense of what you could actually do with AI technology at the time. And what I really wanted to figure out is GPT-3 had a 4,000-token context window, but I really wanted to feed it all of my enterprise data from product PRDs to engineering notes to meeting notes to transcripts to customer conversations. And so I realized you couldn't actually do that out of the box. And so I built this little toolkit that enabled developers to basically load in and connect various sources of enterprise data to an LLM in order to actually get value from it. Synthesize insights, answer questions and take actions over it. So that really formed the foundation for honestly how the company is operated today. That developer toolkit started just as ChatGPT came out and became a very popular framework for a lot of developers across the world to really build their initial AI applications connecting data with LLMs. And since then, we've really evolved as a company to really both provide a comprehensive agentic toolkit to enable developers to build agentic systems on top of state-of-the-art proprietary models today, OpenAI, Anthropic, Gemini, as well as the managed services with building blocks to process and structure massive quantities of unstructured document data and make that accessible to these AI agents. So the company has evolved since then. We raised a seed round series A, we're around 30 people nowadays, and we have pretty big mindshare within the community. Used by many Fortune 500 enterprises both on the open source and also managed services side, and we're really excited about the feature.

Jerry Liu

>> How'd you go from context window, which is a feature that you knew the LLMs were going to eventually catch up to, to being a connector of enterprise data?

Jerry Liu

>> I think it's a pretty natural pathway. If you think about the 4,000-token context window, and even nowadays Gemini has a million, two million context window, there's still a clear need for agents to connect to the right set of tools to actually load in the relevant context if that enterprise data is terabytes or even petabytes large. And so there's a real question of how do you actually inject the right context into these models so that users can get value out of it? And these days with model capabilities and reasoning capabilities getting better, the core need for every enterprise today is not to just depend on the raw reasoning capabilities of these models, but to use it with their own data, and so we provide the right architecture and tooling to make that possible.

Taqi Jaffri

>> Yeah, so just shoving a bunch of data into a context window ...

Jerry Liu

>> Exactly.

>> It doesn't give you context is what you're saying, despite the name.

Rebecca Knight

>> So Taqi, talk a little bit about UiPath's open source strategy here. You are embracing LlamaIndex. What's the philosophy around that?

Taqi Jaffri

>> Yeah. Raghu talked about this on stage in his keynote, so I'll just echo him that UiPath has a open and interoperable strategy and we want people to be able to build agents however they want because we realize there's so much innovation out there that we don't want to hold back from our customers. So we want our customers to be able to build agents using LlamaIndex or any other system that's out there that they feel an affinity to. Now where we add value and where this becomes a better together story is with Maestro. Because as you've seen in this conference, orchestration is our forte. And what that really means is we can bring together all these agents in a single canvas with humans and deterministic systems sort of in this end-to-end automation that's partially deterministic, partially agentic. But then more importantly, we have dashboards over it. We have compliance and traceability and debugging, these kinds of things that you don't really think about when you're doing an initial POC, but really matter when you're taking things to scale.

Rebecca Knight

>> So you have a big developer community with LlamaIndex. Vibe coding is all the rage. Depending on who I talk to, it reminds me of Citizen Developer. You talk to a full-stack developer and they're like, "Come on." Vibe coding, there's kind of a similar dynamic, although it's real. So what is vibe coding, why is it important, and what's your position on it?

Jerry Liu

>> It's a great question and I know that's a multipart question, so I'll try to tackle it one step at a time. Firstly, vibe coding I think was made popular by a tweet from Andrej Karpathy where he basically described the process of entering natural language into a coding IDE like Cursor or Quadcode. You just kind of type stuff in English and then the agent will just generate code for you. And so as a result, you as a developer, you're not really writing the code anymore. You're more just checking the results of what the agent generates. And at first, it's a very fascinating human-agent interaction pattern, but two, it kind of highlights this idea that if you can actually specify English as the default input to actually building these automations, the potential is you actually open it up to an audience that's far beyond developers. And in fact, I would argue the gap between developers and non-developers are closing. We've hired solutions engineers who honestly have not had deep coding experience before, but they're able to build extremely powerful applications with Cursor. We have Fortune 500 customers that are building applications using Quadcode and Cursor, and these are business users and not developers. There's a lot of hype around vibe coding. There's also, of course, a lot of pitfalls. As I'm sure anyone who's tried some of these tools will tell you, they're not super production-ready if you don't actually set the right constraints and I think that's really the trick. If you aren't careful, you end up building a bunch of demoware that's not very useful for other people. But what I'm personally bullish about, and I think this pretty much aligns with what UiPath is going towards too, is if you have the right interface for defining this orchestration, this process automation, with the right set of constraints, governance, and guardrails, you can basically build a natural language interface that translates user and process requirements into the underlying code as the fundamental language for doing this process orchestration. It can be agentic in nature and you can basically kind of create this very constrained yet high accuracy layer for generating various types of automations. And so we're very excited about it because obviously as a developer framework, we have the underlying code-based layer to actually build these deeply specialized agents. And how you actually build the natural language interfaces on top to then translate into the underlying code is an active research area that we're working on.

>> So in your experience, Taqi, what's the best approach to go from demoware to actually production? Because when I talk to BI analysts who they do SQL, they do Python, and now they're actually quite good at it, they'll even use GPT-5 or something to go back and fix their GPT-4 code. So it's getting better, but is the best practice to start with the full stack developer and say, "This is what I want to do," and then the junior analyst goes off and builds, or is it throw stuff over the fence? What are you finding that works?

Taqi Jaffri

>> Yeah, so it's the standard answer, right? It depends. But I'll tell you what we are seeing a lot of which is people start with the business process now. It used to be that they would start with a particular agent or a particular piece of the puzzle. Now with Maestro and things like that, people are starting with the entire end-to-end process, like purchase-to-pay or claims approval or things like that. And then what they do is they initially have a lot of human in the loop, and then what they start looking at is just monitoring that, just using a product like Maestro to just manage how well the humans are doing and debug what's flowing between the humans. And then yes, of course, adding automation where it's clearly the thing to do, and then slowly but surely adding agents as a way to help the humans, draft what the humans are doing, and the humans are still approving it. And then finally, as they can start essentially measuring how much the humans are correcting the agent's output, at a certain point you realize that the humans are barely correcting the agent's output. So at that point, they start dialing back the human supervision. So this sort of starts small. And then "measure your way to success" is what we are seeing is the most successful approach rather than just all agentic, all the way, all the time.

>> So is the subject matter expert, the domain expert, the vibe coder? Is there a middle layer? You'd have a business analyst who would translate tech speak into business speak. Is that middle layer going away? Is it flattening? Does that make sense?

Jerry Liu

>> I do think it's flattening to some degree. I mentioned this on stage too, but the MIT reports where 95% of AI agent applications don't make it to production and the 5% that do. If you look at the report, it basically aligns with our experience too where the 5% that are able to ship to production are very good at deeply understanding the business workflows of the end user and then being able to translate that into an agentic workflow. Whether you do that with a pro code tool like LlamaIndex or kind of like a low code general orchestrator like UiPath Maestro, it doesn't really matter. It's just that you need to have that understanding of both bridging that gap towards business user requirements with the agentic architecture. I think the advantage of vibe code, and it's not quite there yet so I'm going to caveat this a little bit, but the advantage of it is it basically removes a lot of the handoffs you need to basically then do that translation from the product with the technical teams because you as the user can basically almost self-serve build these types of orchestrations without needing deep IT integrations that require a lot of technical effort.

Taqi Jaffri

>> Sorry, I'll just add one very brief thing there. UiPath does have something built into it called Autopilot for developers.

>> Right.

Taqi Jaffri

>> And what that allows you to do is rather than start from a blank canvas and sort of drag in step one, then manually connect it to step two, just describe the process in English and it scaffolds it out for you, and then you can describe any changes to it as well in natural language. So that's a very fluid, nice way of getting started. Now of course, just like vibe coding with actual code, not just low code, at a certain point to take you to production, you really need to sort of fiddle with the details, but it helps you get started very rapidly.

Rebecca Knight

>> So it sounds like the engineers of the future are going to really need to bone up on English Lit classes.

Taqi Jaffri

>> I think so.

Jerry Liu

>> Yeah.

Rebecca Knight

>> But maybe this ...

>> Your dream world.

Rebecca Knight

>> I like that. Maybe this is a provocative or even off-topic question, but what will the developers of the future ... Will they be writing code at all or is this the beginning of the end? And I know you had the big caveat that this is not quite happening yet.

Jerry Liu

>> It's not quite happening yet, but I think it's a very interesting question. And take this with a grain of salt. I might be wrong. I do think you still need a formal language. English isn't quite like a formal language. You need something that can actually verify, can kind of run in a deterministic manner. And so actually, I think the interesting thing about vibe coding is the English to the underlying code piece is more of a translation layer, but how the agent actually runs is through bits basically. It's through this formal language that you can actually compile and run on any machine. I feel like that's still going to be important because people generally want a little bit more determinism in their systems. But assuming vibe coding or this idea of users being able to build automations generally takes off, I basically see a world where instead of this handoff between IT and business process, everybody is able to build their own system of integrations, as well as workflows that works well for them, and that basically we're going to be training everybody to do a little bit of that English-based programming to make sure they speed up their workflows the right way.

Taqi Jaffri

>> I will also just add one thing. I think we are talking a lot about natural language, so English or any other language, to describe the task. There's another powerful thing that's happening which is demonstrations. So we announced at this conference Project Delegate. Other solutions exist out there where rather than describe the task in natural language, you just do it and the agent watches the screen. You hit a record button and then it's like, "Okay, got it. And I will now repeat that task for you." So that's also a very powerful way of programming.

Taqi Jaffri

>> That's such a UiPath thing, isn't it? That's awesome. I love that. When Kubernetes first came out, I think the committers made a really good decision to keep it simple. They didn't try to put a lot of enterprise function in there, and so adoption picked up and then people realized ... It's like that meme. "You don't just deploy Kubernetes." It's more complicated than people realize. So when you think about things like MCP and A2A, how are you seeing the evolution, the connection points, brittle code, versioning, backwards compatibility, patching, things like that? How do you see that playing out and what role do you guys play?

Jerry Liu

>> Yeah, it's a good question from multiple fronts. We have a lot of enterprise customers. We're much smaller than UiPath, but we do have enterprise customers that have a lot of security requirements and so you do need to be careful in terms of the traditional Silicon Valley startup of move fast, break things with making sure you don't actually break your customers' production builds. But at the same time, though, I also think in general, and this relates to the vibe coding comment I just mentioned, foundation models are probably owning the agent reasoning loop a little bit more, you can see this with basically the trend that every model company is going today, which means as a tooling provider, as a startup, a lot of the differentiation is providing high quality tools that can plug into these agents. And we've kind of increasingly realized that from the open source side too. Both the open source and managed services side is we're increasingly focused on how do you actually not try to own all of a agentic reasoning, but provide the right tooling and integrations that provides the right context to these models? And when you think about it from that way, it kind of implies a different process for shipping, as well as being able to deeply handle stuff like authentication, security guardrails, and we're including a lot of that into the managed platform too.

Taqi Jaffri

>> Yeah, I'll just add one quick thing. To us, this question strikes to the heart of why you need orchestration. People just think orchestration is just stitching together different systems and calling it a day. It's not. It's those things you're talking about. What if one of them gets versioned? What if you want to roll back something? What if you want, and I can never use this word correctly, idempotent execution like guaranteed executions, like retry forever? So we use a lot of underlying tech to make all that seamless to the developer and that, to me, is the heart of orchestration, these kinds of things. But to your point, we want to make it simple to the developer. We want to do all this stuff under the hood automatically for them.

Taqi Jaffri

>> Just from a business model standpoint, both UiPath and LlamaIndex, correct me if I'm wrong, but you're taking advantage of the scaling laws. You don't have to worry about the brutal economics associated with scaling laws. You are floating on top of that. It's quite remarkable now that software has become such a terrible business. It used to be so non-capital intensive and now parts of the software business have become ultra-capital intensive. Not your businesses, correct for LlamaIndex? I know that's true for UiPath, but you're not building foundation models.

Jerry Liu

>> It is true.

Taqi Jaffri

>> You're riding off of foundation models.

Jerry Liu

>> Yeah, I think you're right. And it's basically related to the point I mentioned which is we're not really trying to own the underlying model technology or even the agentic reasoning layer. If you kind of more see yourself as providing a really good MCP server and a context provider, it kind of implies different unit economics than actually owning the underlying infrastructure and we do see ourselves playing a key role there.

Rebecca Knight

>> So what has surprised you about the way enterprises are using LlamaIndex? Because open source is so great and innovative, but there are some concerns about its stability.

Jerry Liu

>> Yeah, for sure. Well, I mean, I would say on the open source side, and this is the case for a lot of frameworks at the time, in the beginning we very much were just move fast and just try to keep up to date with all AI advances going on. Since then, I like to think we've matured a little bit. We started kind of creating more stable release patterns, having patch fixes, maintaining backwards compatibility. But it's kind of interesting. As we built more advanced tooling in our multi-agent framework, as well as kind of incorporated it with really good document processing and management capabilities within Llama Cloud, we started seeing a lot of even enterprise customers move beyond building rag chatbots, which is kind of like the first AI use case everybody builds, into multistep, deeply-specialized workflows that actually reflect real business patterns with human in the loop. And so a lot of these are extremely hairy business problems. It's like you need to transform an Excel spreadsheet in a very, very specific manner that an accountant can review and actually produces output for its clients. Or you want to do due diligence on a data room of documents and not just do the extraction, but actually construct the end-to-end financial model. We're starting to see a little bit more of that both with the help of our tooling, the core building blocks we provide, as well as the capabilities of coding agents. It's kind of pretty interesting examples of users actually starting to build real-world, agentic workflows of model business processes.

Rebecca Knight

>> And take that example, Taqi, that Jerry just mentioned. How much of that is built on top of traditional RPA versus sort of greenfield agentic workflows?

Taqi Jaffri

>> Yeah, I think what we are seeing, again, is a mix. So some tasks are just suited more to a non-deterministic agent, like just summarizing a document or extracting something from the document, et cetera. It's really, really hard to write rules for that. So when people think RPA, people generally mean one of two things with that term. Typically, people mean it's a deterministic system. So if I take that interpretation of your question, I think it's pretty obvious for certain use cases when you have to use an agentic system because it's dealing with unstructured and rapidly changing and non-fixed inputs and it requires reasoning, et cetera. But sometimes also, as you know, sometimes people when they say RPA, they mean literally UI automation. It's in our name. And I think that is, in my opinion, going to have a renaissance in the agentic world. Agents can understand screens. We announced Screenplay in this conference, these so-called computer use agents that can literally just look at a computer and click. So I think it's early days with these things, but we feel like our strategy is sound, that no matter where these things go, you will want sort of a deterministic scaffolding around the agentic islands because you need to be able to measure and sort of direct the workflow. And indeed, by the way, we do this with humans today. We call them SOPs or approval processes. So just like we put humans in these sort of deterministic work queues and approval processes and SOPs, we're going to do the same with agents.

Rebecca Knight

>> So what happens to the software stack? A couple of software pros here. Clearly, the infrastructure is changing. It's all going sort of GPU and parallel processing, and that has implications up the stack on the data stack, the BI, players, the whole analytics piece. That middle is sort of fattening. You're bringing in this whole orchestration layer. You're bringing in highly specialized capabilities. It feels like the UI is getting thinner. How do you guys see that evolving?

Jerry Liu

>> Yeah. I mean, to be honest, it's actually an open question. I actually don't completely know the right answer to it, but I do see there's some simplification of the UI layer, but the user does need to actually trust and understand that whatever agentic workflow is being served to them in an AI application is actually context-engineered and that they have the right constraints for input. And so I think a lot of interfaces these days ... Imagine ChatGPT. It's an extremely simple interface, can do a lot of different things. You don't necessarily need a ton of knobs and toggles. But at the same time, there is some advantage in having a little bit of a visual representation of the workflow too so that you have a little bit of transparency into what's going on. ChatGPT as a blank slate also has downsides because you basically have to type everything in order to actually get some results out of it. And so I think it's quite interesting. I do think with this pattern of MCP and agentic reasoning, there's going to be what I call MCP tool providers, folks that basically are data brokers, like context providers for these agents. And then there's going to be these general AI agents, like orchestrator front ends, whether it's a low-code builder like Maestro or it's a simple chat interface or some general automation builder.

Taqi Jaffri

>> Interesting, because that context could come from SaaS. It could come from some emergent system of intelligence with a 4D graph or a graph database.

Jerry Liu

>> Exactly, and those will still be useful. You still need those. It's just I think it's an open question on how much you actually still need the front end for those things.

Jerry Liu

>> Indeed. What would be a real win for you guys in the next 12 months that you would consider, "All right. This relationship, it's working"? What would be a home run for you?

Jerry Liu

>> Oh. Yeah. For the LlamaIndex + UiPath partnership? That's a great question. I think for me, it would be being able to build and deploy a deeply complex document workflow agent, process and insurance claims end-to-end with 99% accuracy using LlamaIndex first in the span of five to 10 minutes, and then also deploy it as part of your overall kind of enterprise-grade set of agents and automations that you have within UiPath also in 10 minutes. I think I have this bar within the company. And we haven't gone there yet because again, we're a pro-code developer tool, just keep that in mind, but how do you build something that's 99% accuracy and you can build it in five to 10 minutes, and then also, of course, integrate and deploy with the rest of the system?

>> Oh, yeah.

Rebecca Knight

>> I love it. The future is bright. Thank you.

Taqi Jaffri

>> Our answer is simple. It's in the title of this conference: ROI. So we measure success by ROI for our customers and we have a lot of our customers here speaking to what they're starting to see. So we are pretty technology-agnostic how they build. We just want them to see value out of it.

>> Show me the money.

Rebecca Knight

>> Perfect note to end on, Taqi. Thank you so much. Jerry, thank you so much for coming on.

Jerry Liu

>> Thank you.

Taqi Jaffri

>> Thank you very much.

Rebecca Knight

>> I'm Rebecca Knight for Dave Vellante. Stay tuned for more of theCUBE's live coverage of UiPath Fusion. You're watching theCUBE, the leader in enterprise tech news and analysis.

>> (Music)