[Emerging Tech] Mastering AI Harnesses and Guardrails With Unmukt Raizada
Unmukt Raizada is the Founder and CEO of TrustEvals, an AI company that helps organizations build and deploy reliable AI with built-in governance, audit, and evaluation. He has 16 years of experience working at the intersection of data and AI. Before TrustEvals, Unmukt held senior data and AI roles at Goldman Sachs and JPMorgan Chase and co-founded the AI software company Thena. He now helps organizations move AI into production with stronger evaluation, traceability, and governance.
Here’s a glimpse of what you’ll learn:
[2:43] Unmukt Raizada discusses how TrustEvals helps subject matter experts turn their knowledge into AI solutions
[4:10] Why AI systems require continuous evaluations rather than one-time testing
[8:50] How AI harnesses guide agents through workflows, prompts, tools, and tasks
[11:22] Why AI agents need guardrails to prevent unintended actions
[14:11] Unmukt talks about balancing probabilistic AI with deterministic systems in high-risk workflows
[20:19] How well-designed harnesses unlock the capabilities of today’s AI models
[26:10] Why observability tools alone don’t constitute effective AI evaluation
[29:27] Using adversarial agents to monitor AI performance and identify problems
In this episode…
As AI systems become more powerful and autonomous, businesses face a growing challenge of ensuring agents behave reliably once deployed in real-world environments. Traditional software testing and basic monitoring may not catch unexpected AI behavior before it creates problems, particularly in sensitive or regulated workflows. How can organizations capture AI’s benefits while maintaining the evaluations, guardrails, and human oversight needed for reliable performance?
Unmukt Raizada, an AI systems and evaluation expert, recommends continuously evaluating AI agents instead of assuming successful testing at launch guarantees future performance. He emphasizes combining probabilistic AI with deterministic engineering where mistakes cannot be tolerated, establishing guardrails around agent behavior, and maintaining human intervention at critical points. Unmukt also encourages subject matter experts to apply their workflow knowledge and intuition when designing AI systems, while avoiding unnecessary complexity from excessive numbers of specialized agents.
In this episode of The Customer Wins, Richard Walker interviews Unmukt Raizada, Founder and CEO of TrustEvals, about building reliable AI systems. Unmukt discusses AI harnesses, continuous evaluations, and the role of human expertise in keeping AI agents effective and controlled.
Resources Mentioned in this episode
"[Governance Series] Transforming Wealth Management Compliance With David Reeve" on The Customer Wins
"Building the Future of Private Markets Technology With Gareth Lewis" on The Customer Wins
"[Governance Series] Revolutionizing Compliance for Small Firms With Paranj Patel" on The Customer Wins
"Building Trust-Driven Tech Partnerships With Brian Hyman" on The Customer Wins
Quotable Moments:
“Building agents is getting easier and easier, but getting it to actually deliver what was promised over time.”
“Building efficiency at the same time, you're reducing your risk.”
“The harness is no longer about sort of controlling the model. It's more about, ‘How do I unlock value?’”
“There needs to be more, more subject matter experts who sort of step up business owners.”
“I think adversarial agents definitely have a place in the solution set. Is it the Holy Grail? No.”
Action Steps:
Continuously evaluate AI systems after deployment: Regular evaluations help identify when AI agents go off script, misuse tools, or produce unexpected results before those issues create larger problems.
Combine AI with deterministic engineering practices: Using AI for probabilistic tasks while relying on traditional software for areas requiring near-zero mistakes can improve efficiency while reducing risk.
Establish clear guardrails for AI agents: Defining boundaries around what agents can access and do helps prevent unintended actions, especially in sensitive, regulated, or privacy-focused environments.
Keep human expertise involved in AI workflows: Subject matter experts can recognize flawed outputs, challenge AI recommendations, and apply intuition developed through years of experience.
Avoid overengineering multi-agent systems: Keeping AI architectures as simple as the problem allows can reduce complicated handoffs, context overload, and difficulties identifying where failures occur.
Sponsor for this episode...
This is brought to you by Quik!
At Quik!, we provide forms automation and management solutions for companies seeking to maximize their potential productivity.
Using our FormXtract API, you can submit your completed forms and get clean, context-rich data that is 99.9% accurate.
Our vision is to become the leading forms automation company by making paperwork the easiest part of every transaction.
Meanwhile, our mission is to help the top firms in the financial industry raise their bottom line by streamlining the customer experience with automated, convenient solutions.
Go to www.quickforms.com to learn more, or contact us with questions at support@quikforms.com.
Powered by Rise25 Podcast Production Company
Episode Transcript:
Intro: 00:02
Welcome to The Customer Wins podcast, where business leaders discuss their secrets and techniques for helping their customers succeed and in turn, grow their business.
Richard Walker: 00:16
Hi, I'm Rich Walker, the host of The Customer Wins, where I talk to business leaders about how they help their customers win and how their focus on customer experience leads to growth. Some of my past guests have included David Reeve of InvestoCOM, Gareth Lewis of Helm, and Paranj Patel of ClearLines Group. Today is a special episode in my series on new and emerging solutions, and today's guest is Unmukt Raizada of TrustEvals. And today's episode is brought to you by Quik!, the leader in enterprise forms processing. When your business relies upon processing forms, don't waste your team's valuable time manually reviewing the forms.
Instead, get Quik! using Quik!. You'll be able to generate completed forms and get back clean, context rich data that reduces manual reviews to only one out of 1000 submissions. Visit quickforms.com to get started. All right, before I introduce today's guest, I want to give a big thank you to Brian Hyman of High Meadow Solutions for introducing me to unmute. Brian is both a strategic partner for Quik! and for our new Quik! for Salesforce app and also a recent guest on my show.
Go check out their website at highmeadowsolutions.com to learn how they help organizations build world class solutions. All right. I've really been looking forward to. Today's guest is the founder and CEO of TrustEvals.ai, where he helps financial firms capture the upside of AI while maintaining strong governance, traceability, and auditability. With 16 years of experience building data and AI systems across Goldman Sachs, J.P. Morgan Chase, Athena AI unmarked works alongside financial service operators to bridge the gap between AI approval and reliable production deployment.
His work spans AI transformation, production evaluations, governance frameworks and audit readiness across banking, capital markets, asset and wealth management, insurance fintech. How about all of them? Just. He's amazing. Welcome to The Customer Wins.
Unmukt Raizada: 02:19
Hey. Glad to be here and excited to speak to you. Rich.
Richard Walker: 02:23
I appreciate you being here today. So if you have not heard my podcast before, I talk to business leaders about what they're doing to help their customers win, how they build and deliver a great customer experience, and the challenges of growing their own company. I want to understand your business a lot better. How does your company help people?
Unmukt Raizada: 02:43
So essentially, we help subject matter experts use their knowledge to build out state of the art AI solutions across, as you said, finance for primarily. So you could be a tax expert who runs a CA firm and is looking to now leverage AI to scale. And that's where we come in to partner with you. You could be a real estate professional who has multiple commercial real estate deals across the states, and you are looking to make sure now that your system is set up in a way where across leases, across CapEx, you sort of have an AI brain. That's where we come in, right?
So think of us as the folks that you call on to take your vision to AI reality. And the reason we are named behind evals is because a core part of our muscle is in making sure that AI works in production by leveraging Evals.
Richard Walker: 03:59
So evals as in evaluations.
Unmukt Raizada: 04:02
Yes. Correct.
Richard Walker: 04:03
Okay. And so what does that actually mean then by looking at evaluations. Is that just testing the system once it's in production.
Unmukt Raizada: 04:10
Yeah. That's a great question. I think the parallel that I draw is sort of the old world versus the current world. So in the old world, whenever you were building software, you would have unit tests and smoke tests and you would have integration tests, then you would have UAT and then you would have production testing. Right.
But when it comes to using AI harnesses, you can't just test at one point in time and then assume that it's going to work for all of eternity, right? So there is a combination of golden harnesses, which is essentially the trajectories that you most expect your AI agent to follow. How often do they follow that? When do they go off script? Why do they go off script?
Did they call a tool they shouldn't be calling? If you had a Slack integration and you said, whenever someone signs up to my website, I want you to send a message on my marketing channel. Are you sure that the message actually went to the marketing channel or did it? Did the agent realize, oh, marketing's not there, but there's a social channel. So let me just put it there.
Right? So those kinds of, you know, scenarios that are muscle built on. And primarily the reason we ended up here is because as part of when we were building for Cloudflare, Etsy, Netflix, Clickhouse, we realized that building agents is getting easier and easier, but getting it to actually deliver what was promised over time is something that requires a deep knowledge of understanding how agents work, where they can go wrong, and how to marry their behavior with your product muscle.
Richard Walker: 06:07
Okay. There's a lot to talk about here, so I want to take it back up a few levels. So first of all, your type of customer has some idea of something they want to do. It's not just, oh, I want to automate my emails. They're looking at the world saying, we have this whole idea of something we want to fully automate, and AI is the solution to do it.
So they come to you. Is that a fair statement?
Unmukt Raizada: 06:27
Yes. That is, I can draw out a couple of examples there just to solidify the understanding. One of our customers basically wanted to build the equivalent of an investment committee memo engine, where essentially in wealth banks and, you know, the investment banks, which are a little more niche, you have multiple stakeholders coming together, and then an analyst that's actually writing up what the investment is going to look like. Why should we do it? What have we identified?
Then you go through the ECC committee, you get feedback. That entire loop was very manual. Now with the expertise of this client of ours, we were able to build out a solution which they're boutique investment banking firm uses to cut down time from six weeks to about 45 minutes. Right. Wow.
Where the analysts basically have all historic ECC meetings at their disposal. The agents have already learned the basis of the analyst where they tend to slip the sector what should be looked at and sort of have a you know very strong back and forth to get them towards a place where the committee can make sense of the proposal much faster.
Richard Walker: 08:02
Wow. You know, AI is disrupting so many different places and disrupting is not necessarily a bad word. Like it's making it better. It's making it faster. It's allowing people to do more of what they're gifted at doing.
I hadn't really thought about investment banking, and back in the 90s, I actually worked in an investment bank, and it was all about word processing, like creating documents and fixing them and editing them and trying to ascertain comparables in the market, etc. you just distilled this down to a super fast mechanism. That's amazing. Okay, there's another topic I want to bring into this because you mentioned it, which is a harness. Now I know what a harness is, but I want to hear your definition of a harness so that everybody who's listening to this can understand when we talk about AI and harnesses, what are we really referring to here?
Unmukt Raizada: 08:50
I mean, I tend to think of it as an end to end sort of flow for what you're envisioning from your agent. Right. So to put it simply, I want you to take an email that comes from my mailbox, and I want you to then understand what's in the PDF, classify it, and then drop it in a particular folder. That's like the steps that you would normally call out. So the harness will then take that and go, okay.
So there will be a prompt which will be a system prompt, and they'll likely be a user prompt. Behind the system prompt is basically saying something along the lines of, you are a workflow orchestrator that's responsible for taking emails and classifying them and the folders into a shared folder. And then you'll have the user prompt, which will basically break down the steps. You'll then have tools. Tools are nothing but okay, I'm supposed to read an email, but how do I do that?
Right. That's where you sort of set up. Okay. You have to use your email tool and the email tool will have its own authentication. So in the old world, you used to have microservices and then you would have these microservices sort of talk to each other through an orchestrator.
In the AI world, you're basically saying a harness is a way to execute a set of tasks by giving it directions, giving it the tools it needs, and the way that it should go about executing it all in the expectation that if everything is followed to the letter, you get the output that you had put in.
Richard Walker: 10:31
All right, I love that. I think that's really elegant. I'm going to give a simplistic version of this, and you feel free to push back on me on this. But I think of just to give people an analogy, if you had a toy racetrack with toy race cars, you can design the track however you want, but the cars have to stay on the track. The harness is the track. You can direct it wherever you want it to go, but you want the cars to stay in the track. And the reason I say it this way is because what happens if the cars go off the track?
And that comes back to what you were just talking about, about how do you manage this behavior? How do you evaluate if they're doing what you want in your harness? If your harness is as simple as prompts, prompts are very flexible, very, very open. If it's as rigid as software constraining it, maybe you don't need as many evaluations, I don't know. I mean, I have my views of this, but what do you think?
Unmukt Raizada: 11:22
Yeah, I think it's a really good analogy. And it works well in terms of where things break. Because just taking that one step further, the pit lane also has a track, but the pit lane for a very specific purpose. Now, if there is a red flag, you need to get into the pit lane. But if there's a yellow flag, you just need to slow down.
And so you need to sort of make sure that when there's a yellow flag, you don't suddenly go into the pit lane and park your car. You might go in to change your tires, but like, everything has a purpose, right? Similarly, if you take like the formula one example, when you exit a pitlane, you can't cross over the line. If you cross over the line, you'll get a penalty because it means you're impeding someone. Now, these are all hard rules that exist across every business that you will incorporate AI on.
But the challenge is, are all these rules being followed to the letter? Right? And I think that is where you will see a lot of gaps, especially if it is a fairly broad ecosystem. So to put it simply, one of our customers basically came to us frightened because they deal a lot with the personal AI brain. And so they had WhatsApp integration and telegram integration and iMessage integration.
And it was very clearly written saying, you can read, but you can't write. And one of the harnesses did end up writing in a test environment, and that completely spooked them because obviously they like privacy is like the number one thing, right? Yeah. And I think you will see that more and more as agents tend to go live in production, the scary bits are the fact that people only find out when it's done, rather than have the guardrails to sort of stop it from going off. So in your racetrack, you always have these tires over really dangerous curves so that if someone goes off, you sort of soften the blow, right?
Richard Walker: 13:29
Keeps them off the track out of the spectators' stands. Yeah.
Unmukt Raizada: 13:33
Yeah, exactly. I think guardrails are meant to be the equivalent of that.
Richard Walker: 13:38
Right? Yeah. I, I don't know if I'm stating the obvious, but you know, AI is probabilistic. It has its own capability to decide which direction to go, how far to go, and it's always amazing to me, no matter how many guardrails you have, it can still go off the track. It can still go outside of the boundaries.
And even if you have evaluations on it, it'll come back and say, hey, we saw it go out of the boundaries. So how do you get to the point where you fully stop this thing, and how do you constrain it so well that it can never leave your boundaries?
Unmukt Raizada: 14:11
Yeah, yeah, I'll give you that exact scenario. Right. So what you are getting AI to do is so critical. And where are you leaning on traditional engineering practices? So as an example, one of our customers deals with highly sensitive compliance and regulatory and tax requirements, right?
So the error margin they're super low. At the same time, the work is so manual that the obvious answer is there has to be some element of AI that can be used here. So what we tend to do there is we tend to use AI to create the plan for how a particular scenario should be executed. Once the plan is created and vetted by a human for that particular scenario, the agent learns it and then executes it using the existing engineering endpoints. The API calls the integrations that already existed, so you get the probabilistic whenever something new turns up.
But whenever it is a learned model, all the agent is doing is making sure everyone did their job right. So building efficiency at the same time, you're reducing your risk. So the way I think about it is where can you afford absolute zero mistakes, right? And if you're looking at a simple AI harness that's going to solve for that. That's the wrong place, right?
However, where are you confident that like, what's the handover between probabilistic and deterministic systems? I think that is the key bit that a lot of organizations tend to get wrong.
Richard Walker: 16:04
Yeah. I personally have the philosophy that software owns the AI, meaning I want software to control everything, and I want AI to do the minimal amount of work, the smallest amount of work necessary. But I want to ask a question about this because, you know, the models keep getting expanded faster, better, more thinking, more rationale, more duration, all sorts of different metrics to evaluate. Are they getting better or not? How important is the model compared to the harness?
Unmukt Raizada: 16:39
I think the models have reached a place now where by itself they like traditionally, the way you would write harnesses would be. With so much context that, you know, the model basically had to figure out very little. If you wanted to get it to work, now it's the exact opposite. Now the models are so well made. Right.
And if you look at the harnesses within Codex or cloud code, it's just ridiculous at this point. How a cursor. Right. Like how you can literally take an idea and go into 15 levels of depth. Run a playground, you know, sort of reward the right agent behavior, have QA and test done and.
You know, go through the entire rigor and build something for you in a couple of hours with very little human input. So now the models are super strong that way. But what hasn't changed or what other. What's gotten more important is because they are super strong. If you don't have the right checks and balances when they go off script, by the time you realize they've gone off script, they've likely gone down a very deep path.
Right? Yes. And so that's where, if you look at it in terms of software engineering terms, you'll see Gary Tan's G stack skills, which a lot of folks tend to use. Again, use smartly, don't use everything blindly, but whether it's the brainstorming skill there or the reviewer skill there, it tends to sort of create these human interventions after asking critical questions. And I think the difference between someone who's really building very good products in AI versus someone who's, you know, sort of churning it out fast.
And you can find that difference very, very quickly is in how often they push pushback to the AI, right? How often do you actually go, that sounds wrong. That's not what you should be doing. How about if you approach it like this, right? And just tying that to this whole conversation in the industry right now of what happens to those experts in, as an example, back end coding, they'll no longer have a job.
They'll be replaced by AI. I actually believe it's quite the opposite, because they'll be the ones who will be able to smell a rat before anyone else and go, that's not what I asked you to do. And that's not the way you should be implementing it because you haven't considered these four scenarios, right? So yeah.
Richard Walker: 19:30
I, I, I've yet to see an AI model create a gut instinct. And it's something that you see if you've had recognition of pattern over a long period of time, if you've been building systems for a long time. I posted about this on LinkedIn. I had this experience where I saw the results that Claude recommended. Here are the options.
I'm like, no, none of these options make sense. Why aren't we talking about another option? Why are we doing this in the first place? And it's like, oh my gosh, you're right. I was so stuck in the details.
I didn't step out and see the big picture again. And that's where the human adds the most. I mean, I think there is incredible value to this process. I'm not going back to the harness and model conversation a little bit further. Are you saying that the models have their own harnesses?
Like you don't need to develop a harness now.
Unmukt Raizada: 20:19
No. What I'm saying is if you look at. So firstly, the reason I got up, was to get the model and the harness together from a clot code or a cursor perspective, just to throw out that really well made harnesses for the most part, are products in themselves, right? Yeah. So code is nothing but a set of harnesses that allows you to literally have superpowers when it comes to building products and codex.
I think the new set of models is just ridiculous, right? Yeah. What I was trying to say is that in certain use cases, people have really figured out how to go deep into harnesses to truly unlock the capabilities that exist in the models today. Coding is a beautiful example, but in other scenarios, they are far, far away from it, right? Yeah.
The more you move to the physical world, the more it's like a complete blue ocean. But even if you go within the digital realm, and I think the new batch of YC will give you that answer, I think there was a hot topic in conversation with Gary there where he tweeted that, you know, about 60 or 70% of the companies were harnesses on top of what you would call traditional businesses, right? So AI harnesses traditional businesses. And that makes complete sense because now the subject matter experts sort of have the empowerment, thanks to AI, to go out there and go, I need to sort of relook at how business should be done, given my expertise. That is where the harness now plays a critical role, right?
The harness is no longer about sort of controlling the model. It's more about how do I unlock value for what I know using the model?
Richard Walker: 22:15
Yeah. I'm going to give a software example just because that's my expertise. I would think that AI has been trained on all software practices. It's read every book, it's read every source code that it can get its hands on. So it's an elite programmer, right?
And in fact, a year ago, I started with that premise. I'm an elite programmer. It's going to write excellent code. No, it writes code, but it's not excellent code. It might work, but it's messy code.
and I realized that you have to constrain it. I always think of harness and constraint in the same kind of capacity. You really need to tell it. I need you to program in this specific style, with this etiquette, with this like thought process, so that you don't just produce code, you produce clean code, you produce really nice code, elegant code that is durable, less complicated, that kind of thing. So I, when I, when I think about where people add the most value with AI, it's their systems thinking, it's their ability to apply what they know to get the AI to act in the way they are, knowing they know how to act.
Do you agree or.
Unmukt Raizada: 23:22
Yeah, I absolutely agree. I think there's a journey that most of the model companies have embarked on because they originally started with a lot of open source, you know, code, and now they've gotten to a place where a lot of the sort of repos, even private repos. Some of them they've got their hands on, and the model training companies have been able to sort of plug a lot of holes. So now the bar has gone up. But at the same time it doesn't discount what you're saying, right?
On the contrary, that's exactly where the next evolution will be. We've sort of got to 80% of coding, and the 80 to 95 is always a journey. And the harder journey, like, you know that, right? As an entrepreneur, right? So I think that journey will happen in parallel, but there are so many sorts of industries that are not even at ten or 20 or 40, right?
Simply because relooking at how you do work and understanding where to use AI is sort of where the taste or the intuition comes in that you were referring to even with coding, right? Yeah. I think coding because it's God's math, the intuition or the art to it is still programmable to an extent. Right? But yeah, outside of that, it's sort of a whole new world.
So I think there needs to be more, more subject matter experts who sort of step up business owners, whether they are in giant consulting firms or in specialized industries, to sort of relook and go, how are we really going to do this, given what models and harnesses today allow us to do, and might not be a 0 to 1 journey in six months, as most people predict, I think it will take a couple of years to really, really nail it for your industry. But no better time to get started than right now.
Richard Walker: 25:33
Yeah, I'm really hoping that my listeners are seeing a demystified version of harness and how do you get things to work, etc. because on the one hand you're like, oh, it's coding. I don't understand coding. On the other hand, it's yeah, but AI doesn't understand what you know. And we got to blend the two together. We've got to bring your expertise, your knowledge of workflows, your knowledge of systems, and get the AI to act in those ways.
I want to go back to the evaluation aspect because, yeah, I mean, anybody can look at error logs and say, oh, there was an error. How do you stop the AI from making the error in the first place?
Unmukt Raizada: 26:10
There's such a beautiful article written by OpenAI the other day in their blog, where they talk about the fact that you can have every observability tool out there, but please don't call that an evaluation. Right.
Richard Walker: 26:27
Right.
Unmukt Raizada: 26:28
And it's, it's unreal to me. The number of times I walk into a room where people go, hey, we built this crazy harness and we've got five machine learning engineers, and they've all got like, you know, masters in systems or they've done their PhD. And the moment you ask them, okay, so how is the evaluation journey going? They go, oh, you know, we are using brain trust and we are using lang fuse and we have a sampling of outputs. And you're like, hang on.
So you have a thousand people using your product and you're okay with just looking at 100 of their interactions. And that gives you confidence. Like surely you thought about it and you realized at some point saying, what am I doing? Like this can't work this way, right? I would build a system and not know how your other 900 customers are doing rich.
You never do that, right? No, no. Right. It's doing and it's that's what's mind boggling. mind boggling to me because I was listening to Uber's.
You know, walk through the other day. They also sample like a company as brilliant as Uber, having all the resources and the ability to pull the strongest AI, some of the stronger AI talent, it samples its AI outputs to figure out how it's doing. So surely I think there's a gap in the industry, and I feel that's a large part due to the fact that we're still treating it like traditional engineering. Traditional engineering would be, okay, let me look at my dashboards. Let me see what's dropped today.
Let me go and figure out why it's happening. And when you have like 5000 customers using your AI product, you can't do that. Yeah, right.
Richard Walker: 28:35
I'm going to put myself on the spot here. You can give me the full critique. And hopefully it's good or and good for the audience. I have a system of 78 agents and that's I mean, I've got this massive harness to control all that. And I've choreographed how these agents work.
They don't all work as a swarm. They work in individual roles. But a couple of my agents, their sole job is to be the hall monitor. That's a high school reference where they are monitoring the other agents' work and they're looking at violations, and they're surfacing those violations, and they have the ability to stop and force an escalation back to me when the violations are severe or at risk. Those aren't that's not the only methodology I have for guardrails.
But I thought, why don't I have an agent whose sole job is to monitor all the other agents and surface violations? What do you think of that approach?
Unmukt Raizada: 29:27
I think adversarial agents definitely have a place in the solution set. Is it the Holy Grail? No. It's the laziest and cheapest. I'm not cheap, per se.
Depending on the model, you go open source, maybe traditionally cheaper. But it's a good way of sort of getting the initial feedback in. So the 76 agents I'll come back to. Right. But in terms of adversarial agents as a thought process.
Absolutely. In fact, on my side, I've written it in a way that whenever we are working through a deep problem, Claude will first give me its version. It will then go and hit Codex. Codex will give me its version, and then Claude will sort of reiterate, so we are doing the same thing. Right.
That's what the adversarial path is. I think what is tending to happen a lot right now, and I see this on like while you were saying it, I could picture three of my current clients because we sort of took them away from that model because they are like, we have ten agents and this agent goes and does this, and this agent goes and does this. And we created it to be atomic. And that's again, it's like a microservice approach, right? Which is this service does this, this service does this, and this service is the checkout.
I think that's creating a lot of complexity because now suddenly you have to worry about handovers, right? You have to worry if one agent gave enough context, but not too much context, because then there'll be context overload. And then the model will tend to go off and do its own thing. And then you have to figure out across those 76 where it dropped, how did it drop? Which trace here means which trace here.
Like I think it's a lot of overengineering. To the extent that the models today, I don't think they're like, I think unless you are building a really, really deep problem. I'd be surprised if five agents are not enough.
Richard Walker: 31:42
I don't know what I've built, so I'll argue it's not enough.
Unmukt Raizada: 31:47
Yeah, absolutely. I'm happy to be wrong there, but I think my point is that we can't look at it in terms of how we looked at traditional software engineering. I know the standard one of the things going on in the industry is keep it modular. I think that's doing yourself a disservice because when it breaks, it's going to be a giant pain to sort of orchestrate it in a way to figure out when, where, how, when should I be notified? How do I make it?
You'll then replace one agent. It will have a knock on impact on five other agents, right? So there's sort of a lot of you need to walk in with your eyes wide open is, I guess what I'm trying to say. Sure. Yeah.
Richard Walker: 32:32
Yeah. We don't have a lot of time to debate this that much. I will say this. The approach that I've taken is very software driven. So the agents that are in use are used for very, very specific purposes.
Kind of like what you said. They have a distinct purpose. And it's not that it's one agent handing off to another agent, to another agent. It's software managing that cycle. So they all play very distinct roles.
They're not all called at once. They're not even all needed for the same pass, if you will. Yeah. But I think there's a lot to this, and I'm glad that we're surfacing this because there's a lot of different ways to approach how you solve your problem. And I 100% agree.
Keep it simple, as simple as possible. The problem I'm trying to solve is really, really complex, and I'll leave it at that. We're running out of time. So before I get to my very last question, I want to ask you, how can people find and connect with you and learn more about what you're doing?
Unmukt Raizada: 33:25
Yeah. So I'm on LinkedIn. Apart from that, you can go and look up TrustEvals.ai. We have a new product called HONE. So there's usehone.dev.
And that's what we are pushing from a reliability standpoint. Outside of that, I think LinkedIn is a great place to catch me. And my email would be the other. Happy to drop that in the link.
Richard Walker: 33:55
Yeah. We'll put it in the show notes for sure. All right. I love this question because it just changes everything about what we just said. Who has had the biggest impact on your leadership style and how you approach your role today?
Unmukt Raizada: 34:10
Yeah, I think I've been super blessed in that my, you know, what people tend to call the inner circle, whether that's family or friends. I think I've had a lot of shining examples from a leadership perspective in my inner circle. So when it comes to the family, both my parents were entrepreneurs, right?
Richard Walker: 34:36
Oh, nice.
Unmukt Raizada: 34:38
And I think for that generation to sort of go out and take those risks sort of taught me to be comfortable with it. And I think a large part of leadership goes to both of them. Right? My partner's been an entrepreneur. So, you know, the support from there is never ending.
And so shout out to, you know, the three of them, I think I'm also a younger kid. So like most younger kids who have an elder sibling to look up to, I've probably got someone who's the strongest person I know. So my inner sort of immediate family is our 90% of the strength that I need for anything, right? And outside of that, I think shout out there to my co-founder, who's again, been an entrepreneur multiple times over, understands the game and knows how to push. And if, as you were asking the question, like to me, it was like, I can name eight people who are all entrepreneurs who are all super close to me, who sort of keep feeding it in, right?
And, you know, happy to name names, but that's like a D to C founder and a restaurant owner and someone who's running a lifestyle business. And it's just, I think I've been really blessed in that instead of 1 or 2 people, my closest circle has everyone sort of going through their own journeys. And almost no one that I know that I'm close to has a traditional job, right?
Richard Walker: 36:29
So I love it. You give me hope that my kids will be entrepreneurs because I'm an entrepreneur.
Unmukt Raizada: 36:38
Absolutely.
Richard Walker: 36:40
All right. I want to give a big thank you to Unmukt Raizada, founder and CEO of TrustEvals.ai for being on this episode of The Customer Wins. Go check out his website at trustevals.ai. And don't forget to check out Quik! at quickforms.com where we make processing forms easier. Hey, I hope you enjoyed this discussion.
We'll click the like button, share this with someone, and subscribe to our channels for future episodes of The Customer Wins. Thank you so much for joining me today.
Unmukt Raizada: 37:07
Thanks a lot. I had a good time. Thanks for listening.
Outro: 37:11
Thanks for listening to The Customer Wins podcast. We'll see you again next time, and be sure to click subscribe to get future episodes.
Comments