GTC Bonus: From Unpredictable to Reliable AI Agents with Ilan Kadar of Plurai

Summary

---

Podcast Summary: GTC Bonus – From Unpredictable to Reliable AI Agents with Ilan Kadar of Plurai

Podcast: Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Episode Date: March 20, 2026
Host: Logan Lawler
Guest: Ilan Kadar, CEO & Co-founder of Plurai

---

Episode Overview

This GTC bonus episode centers on turning AI agents from unpredictable variables into reliable, trustworthy tools for enterprise workflows. Host Logan Lawler speaks with Ilan Kadar of Plurai about how their platform provides a crucial trust layer for AI agents, focusing on simulation-based testing, continuous validation, and automatic guardrails—key innovations helping organizations seamlessly adapt high-performance AI into their operations.

---

Key Discussion Points & Insights

1. Plurai’s Mission: The Trust Layer for AI Agents

• [00:39] Ilan Kadar introduces Plurai as an "infrastructure layer" designed to automate testing, validation, and protection of AI agents, ensuring they’re secure and reliable in production.
  • Quote: “We are building the trust layer for AI agents… to ensure that our agents are fully secure, reliable and they can trust on in production.” (Ilan Kadar, 00:39)

2. The Gap in Agent Testing & Plurai’s Approach

• [01:20] Unlike traditional software, few organizations systematically stress test AI agents before deploying them. Plurai fills this gap by simulating countless user personas and scenarios.
  • Quote: “Think about it like a stress test or a pen testing, but not for security, for quality of the agent.” (Ilan Kadar, 01:51)
• Plurai’s tailored simulators connect directly to agents and automate exhaustive scenario testing for comprehensive quality assurance.

3. Real-World Impact & Case Studies

• [02:49] Ilan details case studies:
  • Financial institution: Had just 50 scam examples before Plurai, lacking confidence to launch. With Plurai’s automated simulation, full test coverage was achieved, enabling safe deployment.
  • Major security company: Faced immediate rollbacks after failed launches; Plurai’s simulations enabled stable, trusted agent performance.
    • Quote: “Once they move to work with Plurai… we automated all the testing process, provide them full visibility and now they are on production.” (Ilan Kadar, 03:13)
• General Insight: The higher the cost of error in a domain, the more critical it becomes to stress test and protect AI agents.

4. Automated Fixes and Guardrails

• [04:17] Beyond just reporting failures, Plurai actively generates corrective steps and automated guardrails—small language models trained to patch vulnerabilities detected during simulation.
  • Quote: “Since we’re generating all the failure points… we are training small language models which are guardrails to protect you against these failure points. Exactly. In production.” (Ilan Kadar, 04:26)
• This shifts validation from a static checklist to a proactive, automated fix-and-guard cycle.

5. Continuous Learning and Adaptation in Production

• [05:30] Plurai departs from a “one-and-done” test philosophy. Their system implements:
  • Ongoing monitoring of agents in production
  • Data drift detection and integration of new data sources
  • Feedback loops that continuously update simulations and guardrails
  • Inspiration from autonomous driving: “Monitor, protect, and then continuously improve.” (Ilan Kadar, 06:06)
• Quote: “Agent as opposed to a traditional software they are always changing… This is the reason why the simulation is continuously running on your system.” (Ilan Kadar, 05:35)

6. Plurai’s Broader Vision & Accessibility

• [06:33] Plurai offers:
  • Research publications and open source resources (Pluri.AI)
  • Demos and integrations for companies to validate and optimize their AI agents for production-readiness

---

Notable Quotes & Memorable Moments

• “We are not only providing testing and recommendation… we are also creating autofixes.”

  • Ilan Kadar highlights the value-add of automated corrective mechanisms beyond passive reporting. (04:17)

• “This is the same cycle [as] autonomous driving – monitor, protect, and then continuously improve.”

  • Ilan draws a direct link between AI agent trust and self-driving car safety engineering. (06:06)

• Host’s endorsement:

  • “I love this. One of the most interesting companies I think I talked to.” (Logan, 06:23)

---

Important Timestamps

• 00:39: Ilan introduces Plurai and its mission
• 01:51: Explanation of agent simulation and stress testing
• 02:49: Case studies: Real-world value of Plurai’s trust layer
• 04:17: Automated fixes and production guardrails
• 05:30: How continuous protection and learning work for live agents
• 06:33: How to find Plurai, resources, and next steps

---

Conclusion

This episode demonstrates how Plurai, led by Ilan Kadar, is reshaping the landscape of AI agent deployment and reliability. With its simulation-driven infrastructure, automatic remediation of vulnerabilities, and real-time adaptation, Plurai addresses key pain points for enterprise AI. The continuous-learning approach—echoing advances in autonomous driving—underscores a new standard for agent integrity in the age of AI-powered workflows.

Learn more at Pluri.ai or find Ilan Kadar and his team for demos and research resources.

---

Transcript

[A]

Welcome to Reshaping workflows with Dell Pro Precision and Nvidia. Where innovation meets real world impact in high performance computing.

[B]

Logan Reshaping the workflows in Nvidia GTC 2026 starting to get a little dire. I'm here with Elon Musk. Oh God, I'm gonna screw it up. You got me all nervous. So Elon's gonna tell you all about it. So tell us a little bit, what is your job? What does Plurai do? And then we'll all jump into some questions. Cool.

[A]

My name is Ila and I am the CEO and co founder of Plurai. What we are building, we are building the trust layer for AI agents and helping companies that are building AI agents in order to fully automate and test their agents in development and to allow and to protect them in production to ensure that our agents are fully secure, reliable and they can trust on in production.

[B]

Okay. I mean obviously agents have become much more popular. We have Clawbot, all types of crazy stuff. So tell me about Plorite. Like where does it set from a security standpoint, right? Like is it something that's setting permissions? Is it something that is rooted a little bit more natively like in your code base? Tell me a little bit more about the nuts and bolts of pluripot.

[A]

We are basically the infrastructure layer that after you develop, you have a prototype, you develop your prototype agent. Now you want to create a test set in order to make sure it's fully tested and to understand exactly how it will perform in production. So this is exactly the infrastructure that we're getting in. We built a tailored specifically simulator that allow us to connect to your agent. And once we define like exactly what is the specification of your agent, we are running a full simulation environment, fully test your agent, creating millions of user and Persona, each one of them trying to stress test a different aspect of your system and then provide you a full report. Think about it like a stress test or a pen testing, but not for security, for quality of the agent. We want to focus and understand that the agent behave as you want him to be. And this is the issue that we are tackling.

[B]

Okay. So it's really like agent modeling in a sense like or simulation. Simulation. I guess a good question to start with is I haven't heard of well Pluri before, but also no one really stress test as their agent. Right. Like at the end of the day, I haven't heard that.

[A]

Right.

[B]

So do you have any stats or figures around someone who's maybe worked with you? You don't have to share company names or anything proprietary. But like where you've seen examples where someone launched an agent in production and then realized it went completely haywire. Like where once they worked with you they were able to show like oh my gosh, this is why this shouldn't have been launched.

[A]

Yeah. So this is actually also what we are showing the demo. One of the financial institutions that has agents for fraud prevention and detection they had before they use us, they have only 50 examples of scams and social engineering attacks. They didn't have enough trust to release the agent to production. They worked on it like for at least one year how to fully test their agent but they didn't have enough trust to test their agent. Once they move to work with Plurai automatically with the click of the button. We automated all the testing process, provide them full visibility and now they are on production and there are many other use cases we are working like in this sense we are not domain specific. So we are working also with one of the major security companies and also for them they saw the pain, they tried to release something and immediately roll back their solution. And we see it like all the time like it's depend, but it's really depend on your domain and your agent. If the cost of error is really high for your domain you must stress as your agent and you must protect it in production. And for that you need the simulation.

[B]

So another follow up question is so I get the stress testing, I get like the validation you want to avoid whose nation and stuff like that. Let's say someone runs like a fluoride test right where it's ran through on the agent. Does it give the corrective steps that need to be adjusted for it to then be validated and to be able to be recommended to put in production. Or is it more just saying hey, this thing is ultimately failed.

[A]

So this is a great question because we are not only providing testing and recommendation what need to be fixed, we are also creating auto fixes. What we are doing since we generating all the failure points and understand exactly what went wrong, we are training small language models which are guardrails in order to protect you against these failure points. Exactly. In production. So we are detecting auto protecting you. And once you are in production we are continuously monitoring the agent behavior and for every edge case that we find, we feed it back into the simulator generating better test better guardrails. So you can think about it as a continuous learning.

[B]

And that one's going to be my next question about the continuous learning bit. You beat Me? No, actually no, you have to be sorry. No, this was not. Yeah, well, I'll ask you. So even though we covered it, but like so you've got the fixes which proactive, which is amazing because a lot of things don't have that. But let's say you have a long running agent that's been running for a year, right? Like how do you. How do you add Explore set in there to have the continuous fixes? Or does it continue optimized because everything every day AI changes a little bit, right? Someone might use the agent a little bit differently. How do you constantly protect the agent from failing if it is in production?

[A]

So this is a good thing agent as opposed to a traditional software they are always changing like also the software itself, like all the agent, all the components of the agent and both the data. There is a lot of data drift. You're adding more documents. Everything is continuously changing. This is the reason why the simulation is continuously running on your system. So in production we continuously monitoring the data, feed it back to the simulator, trying to stress test the new version that you have the new data sources that you added and add a new and updated protection. So it's not like one guardrail, one monitoring. Everything is continuously updated. And this is also we are coming from the autonomous driving space in the previous role we work like with companies like Waymo help them with the self driving system. It's the same cycle, monitor, protect and then continuously improve. Without that agent will be static and roll back.

[B]

I love this. One of the most interesting companies I think I talked to. I really like it. So Elon with Plurai tell everyone website where they can find you if they're interested in learning more.

[A]

So the website is easy Pluri AI. We also have a lot of research paper and open source. You can read about more on that. And you can find me also on the social and Ilan Kadar and I'm the CEO co founder and go to our website blurai AI found our product. Reach us out if you want to book a demo. If you want to start testing and making sure your agent our production ready come to us.

[B]

I love it. So you heard it here. BluRaye is here for your agents AI. Yeah. Thank you. Make it trusted. They trusted and improved and continuously getting better. So with that Logan GTC 2026 I'll see you on the next one. Do what you want. Do what you want. Do what you want. This podcast was produced in partnership with Amaze Media Labs.