Georgie 

Hey Jacky. Thank you for joining In the Blink of AI.

We’re just coming off the back of a lovely four-day weekend, and there’s very few things I would be excited to open my laptop for, but being able to chat to you, one of the most successful Australian AI startups and the founder of Relevance AI. You know, saying that I’m excited to chat to you is definitely an understatement.

Can you please tell the listeners who may not know about you and Relevance, what you are building?

Jacky Koh 

Sure thing. So my name is Jacky. I’m one of the co-founders, co-CEO of Relevance AI. I’m more from a data science and ML background. But more importantly, Relevance AI, we’re building the AI workforce.

So that’s essentially multiple AI agents working together autonomously and collaboratively to help solve and automate increasingly complex and repetitive tasks for the human workforce. We’re a team of around 70 people based in SF and Sydney, and we’ve been fortunate enough to work with some of the largest companies in the world, Fortune 500s, and also fastest-growing scale-ups as well.

Georgie 

When I asked a few of our most successful guests on the show who I had to speak to, Relevance, and you, Jacky, came up almost every single time, you’re such an incredible success. We’re thrilled to have you on the show.

So you build an AI agent workforce for companies and then package it up nicely? Or do you give them the tools to create their own, like DIY agent workforce?

How does it work?

Jacky Koh 

Yeah, it’s mainly the latter. So we are an end-to-end platform that allows people, especially non-technical people, to essentially build AI agents, the tools for agents, and also an AI workforce. So multi-agent systems all in the same platform. And it’s fully hosted, deployed for them.

Yeah, it’s pretty non-technical, no-code friendly. But we also, sometimes, work with customers where we essentially start them off with a template that is already proven, already really good, handcrafted by us or experts like expert users of ours. And that’s also another way people can start with Relevance.

Georgie 

Okay. So it sounds like, you know, you don’t need to have a software engineering degree to work with you guys.

Yeah. Something that popped into my head over the weekend that I added to this question list last night was say I told you, hey, I’m about to go to uni or maybe I’m a high school student about to start uni and I’m going to do a software engineering degree in the world of vibe coding and stuff, do you reckon this is a waste of time now or would you still recommend someone to pursue a career like that?

Jacky Koh 

Yeah, no, that’s actually a really, really good question. So the truth is, like, you know, with vibe coding, there’s a lot of different opinionated ways to solve a problem. And if you don’t know how to code or don’t know at least how the code works, you won’t be able to fully understand the nuances between every kind of opinionated way of building the software.

This means it can cut into costs, speed, etc. Despite vibe coding, and I’m a huge believer in vibe coding becoming more and more important and more and more proliferated, knowing how the actual code works is still really crucial.

It allows you to really get that last mile out of the things you build.

And it’s the same with agents as well, right? You can pretty much get AI to build at least 70% of the agent with our platform, you know, with AI you can essentially type in natural language to build the 70%, but then the last 30%, you still need to get a domain expert, subject matter expert handcrafted, whether it’s through natural language or fixing the dials, prompts, etc.

That’s still really, really crucial. Knowing how these mechanics work behind them is still really crucial.

Georgie 

You know, on one hand, I so agree with you. Like this vibe coding is awesome, right? It is. It’s democratizing getting to that sphere at one stage or non-technical founders, building technical companies, but like, yeah, would I love to… I don’t know if you’ve seen The Matrix, where they put a DVD into someone’s brain and then they have kung fu skills. If I was to pick anything that you could DVD straight into my brain, it would probably be software engineering.

Jacky Koh 

Yeah. I do think, like, you know, going for a degree, and I hope the universities adjust to this, is probably going more and more in depth into software engineering.

Pretty much, you know, you really utilize AI as part of the teaching to say, like, hey, we now know that it’s really easy for someone to pick up coding thanks to AI making it a much more personalised learning experience.

Like I was able to pick up another language and pretty much say, hey, explain this new language to me, purely in Python that I already know, and I could do the same with other, like, more technical stuff. And that’s really personalised learning, right?

So I actually think higher education should go more in depth rather than the level it’s currently going through.

Georgie 

Uh, hot Takes, University lecturers, update your notes please.

Uh, look, let’s get into something I’m really excited to peek behind the scenes of what you guys are building at Relevance. So like, cast your mind back. You met your co-founder in a math classroom, I read about in year eight. Yeah. Why do you guys work so well together? That’s a long history. Are you opposites attract, or are you very similar?

Jacky Koh 

Oh yeah, we’re definitely opposites. Probably in a lot of regards. I think, like, you know, people who work with us, they pretty much know where we are very, you know, clearly opposite. And it’s actually really good because, you know, I’m really aware of my weaknesses and he is as well, and we kind of make up for each other’s weaknesses.

The other parts as well is like, we do share some similarities, probably more so in like, you know, our vision mission, the technological front of things.

So even though there’s differences, it’s not like everything is completely opposite.

Georgie 

How quickly did you realise that and started delegating? Like, Jacky, you are talking to the investors, like where do you guys really see that clear distinction on, okay, this is definitely more of your kind of domain expertise versus the other?

Jacky Koh 

Yeah, it’s actually kind of funny because like, we overlap a lot in terms of what we’re good at technically.

Like, you know, both of us can talk to investors. Both of us know how to write code. Both of us know how to sell and market as well.

But the key thing is like, there’s always like kind of contextual relevance, and at what kind of situation is it best for each one of us?

For example, I’m kind of more well known to be a bit more inconsistent in terms of my IC contributions, whereas he’s a lot more consistent. So, if we need a big pop, then, you know, I’m much more of the forefront person or if we’re in a dire situation, I’m also the person to lean on.

But if it’s more like day-to-day then Dan is better at it. So, you know, there’s more of these personality traits that really complement each other.

Georgie 

Oh, that’s fantastic.

So, there’s an investment committee meeting, or you’ve got to ship product and someone’s sick. Both of you can handle that. It’s all good. You can sell, you can build, you can do all of that stuff. But yeah, the personalities are different. That’s kind of awesome, right? And it sounds like I need to call you in an emergency, Jacky.

Jacky Koh 

Oh yeah, I’m definitely the kind of person, if there is an emergency, I am for whatever reason, very calm and collected about it.

Georgie 

Do you freak out after? Like, are you very calm in the moment and then, then three hours later you’re like, oh my god?

Jacky Koh 

Uh, not really. I think I get into a bit of a zone state, where you kind of don’t realise you’re present. It’s a bit of a weird feeling, but it’s very much like I feel like I’m at peak when there’s a dire situation.

Thankfully there’s not too many dire situations so far. But yeah, it’s definitely one of those personality traits I have.

Georgie 

That’s brilliant. I mean, with a 70-person company, you really do want someone at the head that’s like, not gonna freak out in any crazy scenario. So very happy to hear it.

There’s an article in the Startup Daily, back in December 2021. So years ago, Machine Learning, Data Analysis, Startup, Relevance AI, raised 4 million. So, three and a half years ago, did you expect that you would still be doubling down on agents in 2025? Like can you remember how it felt back in 2021 and the thoughts on agents back then?

Jacky Koh 

Yeah. I mean, I’ll be totally honest. Like, three and a half years ago, agents weren’t really a thing.

People definitely didn’t talk about it. I think it’s closer to three, if not two and a half years ago is when we really pivoted into agents.

But the mission has always been the same. We wanted to essentially bring AI to the human workforce to help automate a lot of the really repetitive, boring manual tasks.

I saw this at my last job at IAG, where there was a lot of people just doing these really repetitive and boring tasks that were quite soul-draining to them.

And when you free up machine learning and actually free up a lot of their time on that task, it makes them happier, it makes them more productive.

They can now focus on things they truly enjoy, and it’s great for the business as well. So, that’s always been the core theme that we wanted to bring to the world.

That hasn’t really changed much. But yeah, in terms of the exact solution and technology we brought to the world, that has changed a little bit.

We pivoted from a vector database company, in fact, we own vector database.com. But we decided that was not the right domain space for us. We were much more passionate about agents and agent automation when we saw that trend starting to happen.

Georgie 

I think the listeners, for the most part, are all over AI agents. You know, it’s the year of the AI agent this year, but what they might not be aware of is how ahead of the curve you guys were. Which is why I’m pumped to ask you, Jacky, what do people get wrong about agents? I remember when we had a quick chat earlier, maybe this month, we talked about single versus multi approach. Please tell us a little bit more about that.

Jacky Koh 

Yeah. There’s a lot of different myths about agents. I think this is probably less of a myth, but more so, when you’re interacting with agents, you might give it multiple different tasks and say, hey, look, it’s not performing that well on these different tasks.

That’s because this agent is trying to act like everyone. And that’s where multi-agent systems come in. You can allow yourself to build agents to be a little more niche down.

So, say, this agent is really good at copywriting. This agent is good at asking questions. This agent is good at synthesising data.

You can break down a longer task into smaller tasks that can be performed by a specific role.

And that will greatly increase the reliability and accuracy of the task being executed.

It also makes everything more modular, so you can reuse these agents you build out. We definitely see multi-agent systems picking up in trend. Google just released a protocol for it. It’s called agent-to-agent (A2A), and it’s helping facilitate multi-agent systems.

Georgie 

I think we’ve done 25 episodes now, and that’s the first time I’ve had that broken down for me in the LLM space. That’s awesome, Jacky.

You did tease us a little bit that people get other things wrong about agents. One more thing that people might get wrong about agents that grind your gears.

Jacky Koh 

Yeah. I think a lot of people, when they talk about agents, specifically in the automation space, they think about traditional automation with LLMs in it.

And that’s the wrong way to think about agents because that’s more just traditional automation enhanced with LLMs.

Like, you know, if you’ve ever used something like IFTTT (If This Then That), that’s a form of traditional automation.

Georgie 

I have used a form of that, but it was, I think, VBA in Microsoft Excel. I used if statements. I know, I know. Look, old school, I’m a millennial. Give me a bit of credit.

Jacky Koh 

Yeah. So, you essentially specify all these rules and then, essentially one of these steps will involve an LLM, and that’s where, you know, people kind of start calling those agents.

Those are in fact not agents. They’re much more just AI-enhanced workflows. So, or workflows with AI in them. So yeah, in fact, you can actually have traditional automation workflows with agents in them, but it’s never like, if I add an LLM, then it’s an agent, so to say.

So yeah, that’s one of the things that people get tripped up about. An agent is a lot more autonomous. You are actually just giving it a whole bunch of tools and you’re not really saying, hey, do this, then that. You do it in natural language in text or prompts rather than like, you know, a flow builder where you’re like saying, if this exact value is equal to this, then you do this.

Georgie 

Oh yeah, we did have a guest on the show that mentioned that the majority of AI agents are glorified workflows, which was the spicy take that went a little bit viral. But thank you for unpacking the if-then statements and the distinction between that and actual autonomous agents going off and completing the tasks.

Look, I’m going to go a little bit left field here, but I can’t not ask you. Tell me the story behind the really cute, kind of Super Mario-like characters on your website, your branding, your t-shirt today, Eddie Will, it’s watching this on YouTube. I’m obsessed with them. Tell us about them.

Jacky Koh 

Yeah. Full context, I am a huge gamer. I played Pokémon when I was young, played other multiple games like Zelda, etc. Yes, always been obsessed with gaming. Still play games when I get the chance to these days.

But yeah, no, the Pixel character really came into mind as something that’s quite nostalgic and also quite relatable in terms of like, hey, virtual workers. We really didn’t want to go down the uncanny valley where we portray agents as humans visually.

So, we tried to find the best kind of alternative and came across pixels. And then yeah, we wanted to go hard on it. It was cute, and it also portrays the meaning quite well. There’s a lot of fun branding stuff you can do with it.

You know, we’re still iterating on it, and yeah, that’s how it came to be. The other part as well is that building agents kind of reminds me a little bit of playing Farmville/idle games where you’re essentially investing time to build it, etc.

But once it’s live, you can log out and come back to collect the fruits of what you built. The agents are working in the background automatically. So, we wanted to send the message that, hey, these agents are truly autonomous on autopilot.

Georgie 

Oh my gosh, that made me want to ask you so many additional follow-up questions, but I’ll try and keep it calm.

Georgina, number one, went to Tokyo at the end of last year. Obsessed. Went to the Pokémon shop, came home with so much Pokémon-related paraphernalia. What’s your favourite Pokémon, Jacky?

Jacky Koh 

Oh, Beon and Salam.

Georgie 

Beautiful. Don’t know either of those, gonna Google them later. Actually, no. I know what they are. I’m into the cute, fluffy ones like, you know, you know the F trick?

Jacky Koh 

F picks, yeah, the Fox.

Georgie 

Fox, and I don’t know why my brain is breaking. I blame the long weekend. Evie. I love Evie. So cute.

Okay, but that’s not a question. My question is, around that uncanny valley, I couldn’t agree more. There’s something a little bit creepy about AI pretending to be a real human.

I don’t know if you’ve played with Sesame, Jacky, and it’s a conversational AI. So, you’re chatting to the AI interface. I find it creepy. I don’t know if you agree.

Jacky Koh 

Yeah, the voice is quite realistic. I think the only thing that, yeah, I find creepy is when they portray it very much as a human visually.

Like, if it was a pixel-style agent talking to me, then it’s just like, oh, hey, these are NPC characters in a game.

But if it’s a clone of myself and it looks exactly like me, I’ll definitely find that creepy. And totally avoid that.

That being said, it’s our opinionated decision. There’s no necessarily right or wrong in terms of branding. We just wanted to go with this brand, and it resonates.

Georgie 

Yeah, I love it. A lot of people love Sesame, but I find it creepy, but there you go.

One of your customers is a household name, Safety Culture. Tell us what it’s like working with Australian startups as customers. Is there a difference you notice, especially in the early stages or in the not publicly listed company stage?

Jacky Koh 

Yeah, I mean, working with Australian startups is honestly amazing.

Like, this is not saying working with US startups isn’t amazing, both are quite equally amazing in their own different ways.

But the great thing about Australian startups is that they’re close, just down the road. You can really utilise your learning opportunity here when teaching them these concepts.

It’s not like SF where everyone’s talking about AI agents, all the latest technology every day. In Australia, people have other topics to talk about. You get more of a chance to enable people, teach them these concepts and verify that they’re widely understandable.

Whereas in SF, people get these concepts more straight away because they have the context.

Georgie 

Yeah, I wanna be in on that WhatsApp group chat for all the Aussie amazing founders. That’d be an epic behind-the-scenes view.

On the same note, you’ve got investors here in Australia and across the pond in Silicon Valley. You’ve got King River Capital here, very prestigious, as well as Insight Partners. Massive.

What’s the difference, like personality-wise, are there any things that are different from the way you speak to your Aussie investors, or are they more similar than different?

Jacky Koh 

Well, they bring really different values, to be honest.

I wouldn’t rate one higher than the other. Our Australian investors, King River Capital, are absolutely amazing. They really help on a personal level.

If I need to have a quick chat or sense check something, I always go to Zab at King River. He’s been really wonderful with that. That’s why we decided to go with them for our Series A.

In terms of Insight, they’re a much larger investment firm, and George has invested in many companies. He’s really great at providing proven advice.

With US investors, you get access to a lot of insights around the industry at the largest scale. Here, it’s more about personal connections with investors.

That being said, King River has invested in multiple unicorns, and we still ask them for advice, too. But US investors typically have more specific industry experience, especially for trends.

Georgie 

Thanks so much for sharing that. In Australia, especially in my role with the AI accelerator for Seed and Series A AI startups, there wasn’t much of a maturity level back then.

It’s really special asking your insights at a later maturity compared to the pre-seed and seed stage startups we see today.

On that note, if you’re an AI founder listening to the show, what are you noticing they’re getting wrong with their tech stack in the early stages? How can they set themselves up better to scale, especially when picking models or tech providers?

Jacky Koh 

Yeah, the number one thing is to not lock yourself into a specific provider.

Model improvements are coming out on a monthly or bi-weekly basis, and the state of the art is changing rapidly. If you lock yourself too much into a provider, you’re making your LLM side inflexible, and you could get eclipsed by competition.

Next week, another provider might be 10x better, or at least 1.5x better. You don’t want that to happen.

So, from the framework front, don’t lock yourself into one provider. Making it work with multiple providers is better, as you can enjoy the benefits of improvements from LLM providers without having to do much besides making it compatible.

When building, what doesn’t work now doesn’t mean it won’t in a month’s time. As long as it looks sound from first principles, there are ways to overcome the last mile. 

Georgie 

I love this. Get those credits, founders and don’t get locked into one place, one domain, one tech company. And yeah, like I hear that switching costs can, like, it’s not that significant, the switching costs, but staying nimble early on, I guess, is smart. Not getting too deep on one infrastructure.

I would love to ask you, Jacky, we’ve never unpacked this on the show before and I can’t think of a better person for this. What’s model context protocol?

Jacky Koh 

Oh yeah, MCP. Um, so I’ve seen this great kind of image that’s a good analogy for this. So think of MCPs as like USB sticks attached to your computer. Your computer is essentially like your LLM.

Let’s say you have a USB stick that is like Google Drive or Zoom, Google Sheets, etc. As you attach these USB sticks into your laptop (your LLM), it now can interact with those different services on your behalf.

For example, if you attach a Google Drive MCP into your LLM, it can automatically upload photos for you, download photos for you, reorganise your folders for you, all with natural language.

The LLM knows how to interact with the different APIs presented in MCP. It allows LLMs to go from just generating text to actually performing actions on your behalf, calling into APIs and working with different systems.

MCP is a protocol that facilitates that, and it’s picking up in popularity.

Georgie 

Do you need any specific prompting to enable it to do that? Like how, you know, you’ve plugged in your USB, then what?

Jacky Koh 

Yeah, to be honest, you don’t actually need that much prompting in terms of giving the task. The majority of the prompting is actually done on the MCP side where you’re explaining what the endpoints in the MCP do.

For example, if I plug in the Google Drive USB, the LLM doesn’t fully know what Google Drive is or how the API works. It probably knows from its data, but if I plug in an MCP for a new technology, like say Apple drops a new API, and it’s not in the data the LLM was trained on, you need to define what these endpoints do so the LLM can understand and interact with it.

MCPs don’t have to be just for popular tools. You can construct an MCP for pretty much anything, but you do need to define what each part does.

Georgie 

Oh, I need to try this. Sounds awesome. Okay, two quick personal questions.

Because, you know, can’t not have you on the show. Before we get to the rapid fire, you told me you’re not much into reading books, Jacky, but you do like articles. What kinds of articles are transformative to you and who do you like to read about to get inspired?

Jacky Koh 

Yeah, that’s true. I don’t read as many books. Bad habit from childhood. But I do love listening to YouTube videos like ColdFusion.

There’s also this one called “Inside Business,” where it explains the business models of different companies. I really love watching those. They’re quick, but sometimes go deep, like hour-long videos.

I also read tech articles. I’m on Twitter too much as well, getting all sorts of information. Paul Graham has some great essays on his website.

One leadership principle I love, which came from Twitter, is by the CEO of MongoDB. He said that to be a great leader, you need to be good at three things:

1. Be a great individual contributor (IC). You need to know the task you’re leading.

2. Be a good manager, knowing how to motivate and support people.

3. Know how to grow the team.

It came from a short form video, but it resonated with me a lot.

Georgie 

You’re not joining a book club anytime soon?

Jacky Koh 

No, I’m not big on reading physical books. I try audiobooks, but I prefer condensed versions of things.

Georgie 

And you’re a nonfiction kind of guy. You’re not reading about dragons and fairies?

Jacky Koh 

No, no. I’m more of a TV person. When people say the book is better than the TV version, I’ll never get to experience that because I’m not going to read that book.

Georgie 

I love how much you know who you are. You’re like, yeah, that might be true, but can’t be bothered. Sorry, I’m out.

Whereas I’m the opposite. I’ve got a five-year-old and a three-year-old and I’m already thinking they cannot watch Harry Potter until they read the books.

Jacky Koh 

It’s not that I think it’s worse, I just know that I’m not the kind of person to sit down and read a book.

Georgie 

100%. Even socially, I’ll hang out with friends, and some love the podcast, but others are like, I’m never going to listen to a podcast. Not because I don’t like you, but just not into that format. You can’t force someone to absorb information in a way that doesn’t feel natural.

Jacky Koh 

Exactly. It’s about knowing what works for you.

Georgie 

Last question before the rapid fire. We had Will Liang, the CTO of MA Financial, on the show. He said he was a “purple person,” which means he’s in the middle between business (blue) and tech (red). You mentioned you’re a T-shaped person. What’s a T-shaped person?

Jacky Koh 

A T-shaped person has picked a vertical they really care about and excel at, but they also expand their skills horizontally.

It’s great for two reasons: You have an area you’re confident in, and you can learn from other subjects that complement it.

For example, neuroscience and machine learning. People in ML draw inspiration from neuroscience to better understand how the brain works, which helps in machine learning.

Similarly, marketing and machine learning both rely on A/B testing. Understanding this from marketing helps you in ML as well.

Being T-shaped means you know one thing deeply but have broad knowledge in other areas that help you improve in your main field.

Georgie 

Oh, you’ve made me realise something about myself in the last six months since I started focusing more on communication. Even though my background is technical, I’m actually best at people stuff.

Jacky Koh 

Exactly. And now, you can see how horizontal skills improve your main focus. Whether it’s engineering or finance, it all complements the thing you really enjoy.

Georgie 

Yeah, and the empathy when communicating with others is so much stronger when you’ve experienced what they do.

Georgie 

Rapid fire questions time. You ready for some spicier questions to finish the show?

Jacky Koh 

Sounds good, let’s do it!

Georgie 

People glorify Silicon Valley, Jacky. But what sucks about it?

Jacky Koh 

It’s not the greatest place to live in. It’s not clean or affordable, but the talent density is world-class. Living there sucks, but the people make it worthwhile.

Georgie 

It’s hard coming from Australia to anywhere else. I went from Australia to London, and I was like, “What do you mean there’s no beach?”

There was no mention of AI in the Australian federal budget. Why do you think that was, Jacky?

Jacky Koh 

I think it’s because they’re not as on top of it. The average Australian probably doesn’t think about AI much. So they didn’t include it. Once they have more knowledge, they might introduce something.

Georgie 

That’s such an intelligent answer. It’s a point I hadn’t considered.

What’s one scary headline we might expect about AI this year or next?

Jacky Koh 

A country could try to completely ban AI. It sounds like a Black Swan event, but I could see it happening.

Georgie 

How do you keep agents ethical?

Jacky Koh 

It’s about the use cases we target. We focus on ethical use cases and have guardrails in place. We avoid use cases that could cause harm.

Georgie 

If you’re an unethical customer, don’t bother calling Jacky.

Georgie 

How are large incumbents like Microsoft and Apple going to struggle in the world of agents where nimble companies like Relevance don’t?

Jacky Koh 

It’s all about pace. Large companies spread their best people across different initiatives. They can’t move as fast as smaller, focused companies. They also struggle with the flexibility of not being vendor agnostic.

Georgie 

Spoken like a true leader, Jacky. Before you go, anything you want to shout out to the listeners?

Jacky Koh 

Shout out to Relevance. We’re going to keep making our platform better. We’re not stopping until we’re number one!

Georgie 

What about the merch, Jacky? When can we buy Relevance merch?

Jacky Koh 

It’s a more frequent question than expected. Hopefully soon!

Georgie 

Thanks, Jacky. Have a great day!

Jacky Koh 

You too. Thanks!