SonicJobs CEO Mikhil Raja on Using AI Agents to Connect the Internet, starting with Jobs - Ep. 233
27:33 Share

Hello, and welcome to the NVIDIA AI Podcast. I'm your host, Noah Kravitz. There's been a lot of talk over the past year or two about whether or not AI will take jobs away from humans. Our guest today, however, is already using AI to connect more humans to more jobs, which is good for job seekers and good for employers.

Companies in the United States alone spend $15 billion annually on clicks to advertise their job vacancies. But 95% of all job applications are abandoned before they're completed. That's a big problem for a lot of job seekers and a lot of employers. Mikhail Rajay is the co-founder and CEO of Sonic Jobs, a startup and a member of NVIDIA's Inception program for startups.

that's making it easier for more candidates to apply to more jobs. Sonic Jobs uses a unique approach to agentic AI and web agents, which I'm excited to dig into with Mikhail now. So let's welcome him on. Mikhail Raja, welcome, and thank you so much for joining the NVIDIA AI Podcast. Thanks, Matt, for having me. So there was kind of a lot in there in the intro, kind of teasing about Sonic Jobs, but since we have you, why don't you tell the audience a bit about what Sonic Jobs is?

Yeah. So companies in the US spend $15 billion on advertising their vacancies. And the way they do that has moved from per listing, which is how they used to advertise, to per click, where

where it's performance-based. And the most important metric in that performance-based recruitment advertising space is the conversion from the paid click to the completed apply. And as you touched on in your introduction, industry-wide today, only 5% of

people complete the job application, which means 95% of them abandon that application. That's astounding. Only 5% of applications that get started are actually completed and submitted. That's, I mean, I figured, yeah, I know applications get abandoned, but that is amazing. Forgive me. I just wanted to express that. Yeah.

Yeah. And the biggest reason for that is the redirection. So candidates today still go through the same 1.0 experience as you did in the 90s, where when you want to apply for the job, you start on the job platform. And then for each job that you want to apply for, you get redirected to the company site to apply. And as soon as you're redirected,

you have a 70% bounce rate. So the majority, the biggest factor in that 95% is this redirection step

which causes the huge friction that we experience in the market. Right. So just so I understand from the job seekers perspective, I'm on whatever, I don't even know, Dice or LinkedIn or Monster, Indeed, whatever they are, the job boards. I see a job for podcaster at NVIDIA and I click, I'm making this all up, I don't know, but I click. And so then the redirect is it takes me from the job board to NVIDIA.

in this case, NVIDIA, but whoever the employer is site. And you said you're losing 70% of people just on the bounce. Exactly right. Yes. Okay. And historically, people have tried to build APIs to connect this whole ecosystem, but it

it's too fragmented an ecosystem. So there's hundreds of job platforms, some of which you touched on just now. There's over 200,000 employers that advertise vacancies in America. And there's over 10 million jobs, each of which will have different flows and different requirements for each application process. And so

APIs haven't worked, which is why you still have this experience that we touched on, which is actually still the 1.0 experience, web 1.0 experience, where you start and then you're redirected and you have to create the application. So the idea of the API would be like if I was on a job board where I'd already entered some info or maybe they have a profile of me and it's got my basic info and I

The API theoretically would take that info. And as I'm redirected to the employer site, it would carry some of that with me. So I don't have to reenter all the basic stuff that you always have to reenter. Got it. But those APIs don't work. I'm seeing you nod. I'm thinking, well, the whole web is made out of spaghetti at this point, but I'm not a developer. So is that what's going on with the APIs?

Yeah, exactly. It's not worked. And just to elaborate on your point, ideally, you would apply directly on the platform that you're on. When you're on booking.com, you don't get redirected to Hilton. If you want to book a room, you can just book on booking.com. And so if you take that analogy here, you should be able to apply directly on the job platform.

which is exactly what we've built. So instead of using APIs, what we use is AI agents, which I know you want to talk about. I'm happy to go deeper into that, where effectively we understand using computer vision and also the HTML, every single input field of the job application so that when the candidate's on the job site and presses apply, we ask all of the relevant questions directly on the site

And then the AI agent takes all of that data and on behalf of the candidate submits their data onto the job application flow to complete the application. And so instead of that 5%, which we touched on earlier, and that 95% wastage, we've got a 26% conversion. So five times better than a standard job application flow, which is fantastic.

It is fivefold increase. I want to divert from my notes here and ask you about that business end of things, because I'll just ask you this question, then we'll come back to the business stuff. But how long did it take you to achieve that fivefold increase in performance over the industry standard? Yeah, we've been building our technology for five years, and it combines what we call traditional AI with technology.

and kind of the new wave of AIs. Yeah, happy to talk more about that as well. Yeah, yeah. So pin in that bad question for me. So it's a teaser for later. Stick around, audience. We'll find out about how Sonic Johnson has grown so quickly. But you mentioned computer vision and you mentioned looking at the HTML. So maybe let's dig in. I sort of want to ask you from two perspectives first.

From the user perspective, I just want to clarify kind of the sequence and where the agent comes in and starts working. And then I'm curious about how the computer vision works and then also what's happening. I think you said HTML behind the scenes. Yeah. So from the user perspective, they stay on site. So you get this win-win scenario where the candidates answering all the questions, they stay on the site that they're on, you know,

the good analogy is the booking.com experience. You fill out all of the details on the platform itself. On the dashboard. Yeah, exactly. Okay. Exactly. And then you're done. You press supply. It's all done. From there,

The employer's perspective and kind of how that all maps through, because we're using AI agents and the input fields are all publicly facing from a job application flow perspective, they don't need to do any development work.

So we're asking exactly the questions that are needed for that job apply flow that's on their website. And they don't need to kind of build an API or have tech resources, etc. And they only pay when they, the employer, etc. Exactly. And they only pay when they receive the application. So they pay at the outcome. Right, that conversion performance model, yeah.

Right. So is the CV, is your agent sort of looking at the job application webpage and visually making sense of what's being asked? Like, oh, this is the first name form, this is the last name form, this is the, why do you want to work at...

you know, Acme company, SA field. Like, is that what the CV is doing? You said CV, but I think you meant... The computer vision element. It jumped out to me when you said computer vision before, so... Yeah, CV in the UK means... Oh, right. Sorry. So I got... Of course, my bad. So yeah, that's exactly what the agent's doing. The average job application form has six pages and 40 input fields.

And so the agent is understanding every single input field. It can be radio buttons, dropdowns, open questions, open text, conditional questions, all that sort of stuff is transforming that to kind of a context and to a structure and then asking those questions exactly right. And also inputting that data back into those questions.

So maybe that can lead us into this talk about AI agents sort of broadly. Because I think a lot of what I'm thinking about goes back to like a year, even a year and a half ago, where there was buzz around auto GPT and some other open source projects.

Talk to us about agents, how Sonic Jobs uses them, maybe how that's different from other approaches. Yeah. So you're exactly right now. You know, 12 months ago, AutoGPT, Baby AGI. Baby AGI, yeah, yeah. Captured our imagination. It was like, you know, the agent that can do everything. Yeah. What we've seen...

is that the architecture of the agent needs to combine very tightly the application layer and the reasoning layer. And that's because accuracy and reliability are really important for companies. You know, the challenge with...

you know, baby AGI and auto GPT was that LLMs are not deterministic. And so they can't do step one, step two, step three, step four in the order that you want them to whilst an enterprise client wants...

the output and the workflow to be structured in a particular manner, which is domain specific. You know, they'll have specific steps to, and again, our domain is obviously job applications. And that means that it's really important to combine

in our experience and in our view, to combine what we think of as traditional AI, which is very high on accuracy and lower on, let's say, generalizability. Combining that with DLLMs, which can actually generate data and use the memory of the successful training data to basically

use that for future use cases, for future workflows, et cetera. And so you've kind of got this combination which we think is really powerful and very domain specific. And that's where our experience of kind of this vertical AI agent

is that it's much more powerful for the specific B2B use cases versus this general agent that can kind of do everything, which actually ends up doing nothing rather than being able to do everything, which is kind of interesting. This is me we're talking about. I'm not a developer, but my own experiments with those early systems did a whole lot of nothing as it turned out, but that's a separate take.

So this approach to agents that you're talking about, when Sonic Jobs, well, at whatever point in the Sonic Jobs story, did you sort of see this approach to agentic AI and kind of think, oh, this maybe is a good solution to this problem we're seeing with the redirects and the other things in the recruiting industry? Or, you know, did it come about another way? What was kind of the moment where you thought, oh, this could be the approach we use?

Yeah, it's a really good question. I spent three years before Sonic Jobs at AutoTrader, which is a car market place. And I became obsessed with this idea that you could remove friction and have much better engagement by users.

And particularly as mobile came in, what I saw in the job space was that this redirection point of friction was becoming a bigger and bigger issue. And we thought about it in the context of really creating this API-less API. And so we create kind of this version of an API where actually the company did no tech work, but...

the candidate could have a really seamless experience like they do on booking.com. And yeah, it's, it's taken us the majority of our existence to build what we've built. So, uh, to answer your question, you know, we're 28 people, 24 engineers, four commercial people in, in, uh,

myself. And of the five years, we've spent over three years just building out the technology. And last year, we launched in the US. That's gone super well. And we're probably

primarily a US business now. But yeah, the engineering part of and the agent part of our business is at the very heart of what we do. Of course. We were talking before we started recording that you moved from England to the US. Is the rest of the company global in the US and England? Where are the Sonic Jobs people? Yeah, we're a fully remote team.

I should mention also my co-founder has a background in robotics and AI. So that's how we all kind of came to this conclusion. And at the time we started, this didn't even have a name. It's amazing that everything's called AI agents now. It's a term and agents kind of represent lots of things now. Maybe we could call what we do AI web agents, specifically for the web. And we used to call it kind of AI plus RPA agents.

No one had a clue what that meant. What's RPA? Robotic Process Automation. So it's kind of workflow automation, which is a little bit more manual. And so we were kind of combining the two. But yeah, the agent term has taken off and it's come a long way in the last five years. I want to ask you more about the agent thing. I want to go back for a second, though, to the company. You mentioned that

the business has sort of taken off and shifted focus to the U.S. So you're primarily working with U.S.-based employers? Yeah, we work with some of the largest U.S. employers today. So, you know, DISH, CVS, Walgreens. Fantastic. And we've seen huge traction for our use case.

And then also mentioned in the intro that you're in the NVIDIA Inception program. Yep. NVIDIA has been awesome with us, and we're excited. We're building kind of deeper RAG infrastructure, all with NVIDIA's help, and it's been, yeah, great collaboration. What's on the Sonic Jobs roadmap right now? Is it just businesses booming and you're doing your thing? Is there a product roadmap that you're oriented towards?

Yeah, I think it's really important to emphasize how early we are in this journey on agents in particular. And, you know, it's taken us five years, but we're still, I think as the whole ecosystem, still at the very beginning.

start of that journey. And so today, our agent can interact with about 60,000 jobs in the US. So understands the input fields from 60,000 jobs. The total market size just in the US alone is 10 million jobs.

And so there's a long way to go in terms of even with our specific domain, there's a long way to go in terms of being able to understand every single input field, every single job,

Is it a matter of just continuing to chip away at a sort of very big, unwieldy problem, which is what you described earlier about making sense of the multi-page application form? Or are there other technical things? We're talking agents on the pod. So, you know, if there's more interesting stuff you can talk about, I would love to hear. Yeah, no, I love this. I love this. The real answer is we don't fully know. There is...

a scale element to this where the more that you get successful applications going through and we've got you know now millions of successful applications that have gone through our agent the more the model learns and is able to adapt to the next

job application flow that it sees, which has been very successful as a technique and continues to bear fruit for us the more clients we work with, etc. There's a reasoning, a general reasoning layer. So where you think about the kind of LLMs are not great, as we talked about earlier, following steps, planning, etc.,

I'm confident that will get better as well. And that will enhance our kind of vertical approach within the application layer as well. I think it'll be a combination of a few things that basically end up kind of being able to

create a vertical AI agent across the entire domain. You're talking about the application layer and the reasoning layer. And if I understand it, or to the extent I understand it, it makes a lot of sense to me that you've got, whether it's processing job applications or submitting, I should say, I guess, or another business enterprise task, whatever it might be. So I understand that sort of the verticalized specific

tasks that the agent is being developed to do. And then underneath that, you've got this reasoning layer you talked about, which is the LLM. Yeah. And so when you're talking about the reasoning layer needing to get better, correct me if I'm wrong, but my understanding is you're not building foundational models from scratch, but you are doing a lot of work to kind of fine-tuning them is the right way to say it.

Exactly right. To get them to perform how you want them to in conjunction with your system. So all this is getting to my question, which is, is there a lot that a company like yours can do to kind of tweak the reasoning layer? Or is it mainly a matter of waiting for the various companies of the world who are releasing foundational models? Is it kind of more like waiting for them to ship something better and then you can...

Or is it kind of a combo of both? No, there's a ton you can do on the fine-tuning side. And also what we call on the tooling side. So traditional AI now is referred to almost as tools. And the LLMs can combine with these tools and the fine-tuning that you put on top of the LLMs to create this kind of hybrid structure. And you touched on it then. It's an important point to emphasize that

In domain-specific workflows, you don't want randomized outputs. You want predictable, accurate outputs. And so depending on your domain, you need to structure the architecture so that you might even choose to kind of hard code or, you know,

create specific layers that do specific tasks and then specific layers that create, that do reasoning tasks or do error resolution tasks or do input detection tasks. And so our view, and you asked about kind of all developers, so I'll maybe touch on that. Our view is that there are always going to be application layer architecture that's going to be needed to create

a vertical specific use case for a particular enterprise. And that's going to be always really valuable. And again, you're kind of riding on the wave of the LLMs becoming smarter because then you need to do your value add can be more and more and you can scale faster and faster. But there's a ton you can do even as a small company like ourselves.

to create that domain specificity, which is hugely valuable. So your technology, and not that Sonic Jobs is focused on anything besides the

talent acquisition industry. But is the approach you're taking, could it serve as a framework, even just conceptually, for, you know, kind of similar work in a different domain? Yeah, I would answer your question in two ways. One is we think that

you know, as you get better at a specific domain and vertical, it actually gives you the springboard to potentially explore other domains, which we think is interesting in a way that we think is more valuable and more likely to be more successful than starting as a generalist and moving to being a generalist, actually starting at a vertical and then moving out to other verticals we think could be more fruitful. The second thing is, and it's a little bit of a

As I mentioned, I'm new to Silicon Valley, so I'll say how I see it, obviously, as someone new. I think AI and AI agents today have been largely built by people who are excited about the technology and looking for a problem.

I think there's a lot of room and I think the other group is the group of people that have a problem, whether that's a B2C problem or a B2B problem, and then look at AI as creating a solution that could never be created before.

I think the way we've done it is the latter in that we had a problem and we looked at AI to create a unique use case. And I think I would encourage anyone listening to this. And you mentioned that, you know, there's a lot of engineers that listen to this. I would encourage

encourage anyone with a problem to know that it's open and not to... Often it can feel like a wall of God and AI coming in from the outside, particularly if you're not from Silicon Valley. I would encourage people to lean into the problems that they have. I mean, even at my sort of 50,000 foot level, what you said just kind of rang true. Simple as it sounds, that idea of LLMs in particular...

All of by nature are this semi black box or just this, you know, emerging capabilities, all these words that are used to describe this kind of general process of figuring out what it can do. And so that approach of starting with I don't know, I personally, if you give me a blank sheet of paper, I have the hardest time coming up with words. But if you say, hey, here's what we need. So, yeah, there's kind of a similar thing resonated with me hearing you say that.

All right. Got one last question for you before we wrap up. It's a little bit of a flip. We usually like to ask how the work you're doing and AI's impact on the work you're doing is going to affect things going forward, which we've talked about a little bit already.

But since you're in the hiring space, so to speak, and we've been talking about the inefficiency of the posting model, changing to conversions and technology like Sonic Jobs, increasing that conversion rate. Again, to put you on a spot here,

What advice would you give to a company who's trying to hire right now? You know, depending on the industry, we're hearing a lot about, you know, the economy and employment in certain industries, talent crunches and others, you know, can't find a qualified person. So kind of given that broad spectrum, what advice would you give to a company who's trying to hire right now?

Are there any kind of high-level pieces of advice you'd give to somebody out there who's trying to advertise and hire for roles but just having a hard time? Yeah, I think the simplest piece of advice would be to go to a job platform

find your job and apply for your job on the job platform. What most companies do is apply for their job or look at their apply flow on their own company site. Oh, but not from the, yeah, okay. But you've got to remember that most candidates start their journey on a job platform, whether that's LinkedIn or Indeed or Sonic Jobs, and really think about that experience from the job seeker's perspective because that's

If you do that, you'll optimize your own conversion and have a better experience for your candidates, which is fruitful. For folks who are listening and would like to learn more about Sonic Jobs from whatever perspective, the technical perspective, maybe they're an employer who's looking for new ways to hire, or I guess maybe job seekers can go look at job listings as well. Where would you send people to go?

online, what all is on the Sonic Jobs website, and then is there social media as well that listeners can follow? Yeah, so if you're an employer and you want to talk more, add me on LinkedIn, Mikhail Raja. If you're a job seeker, AI engineer,

that think we're doing something cool, we're hiring. So please reach out to me. And if you just want to learn about our technology, we've created a page, sonicjobs.com.ai agent on our website where you can learn more about our technology. Great. Sonicjobs.com. Exactly. We're into the age. It used to be way back when it was .com, but people would use other ones because they couldn't get the name. And then it was all .com, you know, org.

Now we're getting like the .ai and the other, so I try to remember to ask. So sonicjobs.com, fantastic. Mikhail, I feel like we could talk agents and talk the future of the web, you know, for longer, let alone the talent acquisition industry. But I think that this gives a great purview into what Sonic Jobs is doing, has been doing, and also kind of bringing back up the topic of agents today.

which is interesting to think about how things have changed in a short amount of time. But it feels like a long amount of time since Gen AI hit the scene. All the best to you and your whole team. Thank you for coming on. And yeah, any words in closing? No, thanks very much for having me. Really exciting. Fantastic.