Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
How Madhav Krishna’s Vahan.ai Became India's Largest Blue-Collar Hiring Platform by Using AI Agents
0 Plays2 seconds ago
Madhav Krishna's journey from Silicon Valley to solving India's 450-million blue-collar workforce challenge is a masterclass in finding product-market fit. Starting with a voice-based English teacher called Lakshmi in 2016, Madhav pivoted three times before discovering the real painkiller: recruitment, not training. Today, Vahan.ai powers hiring for Zomato, Swiggy, Blinkit, and Uber using a unique agency-powered model combined with GPT-4o voice AI that costs just ₹2 per minute compared to ₹3-4 for human recruiters.
He shared the complete journey in this candid conversation with host Akshay Datt, revealing why his WhatsApp bot with 500K users meant nothing until he embraced local recruitment agencies instead of trying to disintermediate them. From Y Combinator to backing by Khosla Ventures and Temasek, Madhav explains why SaaS fails in India, how outcome-based pricing became their moat, and why India's trust deficit requires human intermediaries even in the age of AI. With 150 employees generating 40,000 monthly placements and a clear path to profitability at just 3X scale, this is essential viewing for anyone building in India's gig economy, exploring vertical AI applications, or trying to understand what actually works in Indian enterprise markets.
Key Highlights:
👉How Madhav Krishna built Vahan.ai from ed-tech pivot to India's largest blue-collar hiring platform with 40,000 monthly placements
👉Why 500K monthly active users meant nothing and how the agency model became the breakthrough to true product-market fit
👉The contrarian insight: why engagement doesn't equal revenue in India and how outcome-based pricing beat SaaS subscriptions
👉Building Voice AI for India: how GPT-4o, proprietary call data, and Hinglish capabilities created a ₹2/minute recruiter cheaper than humans
👉Lessons from scaling through India's funding winter: the path to EBITDA profitability with just 150 employees and capital-efficient growth
👉Why blue-collar workers don't look for jobs online and how Vahan.ai digitized 2,000 local agencies instead of disrupting them
👉The future of gig economy hiring: AI agents, regulatory changes, expansion to manufacturing, and the vision to reach 1 billion people globally
#MadhavKrishna #VahanAI #BlueCollarHiring #GigEconomyIndia #VoiceAI #AIRecruitment #IndiaStartups #YCombinator #ProductMarketFit #QuickCommerce #ZomatoSwiggy #DeliveryJobs #GPT4o #VerticalAI #OutcomeBasedPricing #SaaSIndia #KhoslaVentures #StartupPivot #IndiaLabourMarket #WorkforceSolutions #HRTech #RecruitmentPlatform #IndianGigWorkers #AIForIndia #StartupFunding #FounderJourney #TechInIndia #BlueCollarJobs #HiringPlatform #FutureOfWork
Recommended
Transcript
The Skills Gap in India's Education System
00:00:00
Speaker
We produce 7 to 8 million graduates a year, but more than half are unemployable. Training ultimately is really nice to have. It's a vitamin in a way, whereas the real painkiller for these companies was to hire people.
Vahan.ai: Revolutionizing Blue-Collar Hiring
00:00:14
Speaker
Madhav Krishna is the founder of Vahan.ai.
00:00:18
Speaker
Vahan is India's largest blue-collar hiring platform that has placed over 1 million workers in companies like Zomato, Swiggy and Uber using AI.
Challenges of Business Models in India
00:00:27
Speaker
The kind of default playbook in the Valley is, hey, you you build engagement and then that will translate into revenue.
00:00:34
Speaker
ah That usually works in developed countries. That doesn't work in India.
Trust in the Indian Market
00:00:37
Speaker
People in India are time rich, but money poor. Essentially, like India is a low trust market. Therefore, you need an agent of trust. Our current voice bot costs close to two rupees per connected minute, which is in fact cheaper than a human.
Vahan's Cost-Effective AI Innovations
00:00:50
Speaker
What makes it hard for someone else to come use the latest ChatGPT model, create a voice bot and replicate what you are already doing?
Madhav Krishna: A Journey in AI
00:01:08
Speaker
Madhav, welcome to the Founder Thesis podcast. You've been in the AI space for more than a decade. ah Take me through your journey. like like I believe you ah it started with an education in the US to learn AI. How did you get interested in AI and what was the state of AI back then?
00:01:28
Speaker
um Thanks for having me, Akshay. It's a pleasure to talk to you today. I'll tell you a little bit about my background. um I'm a developer software engineer by training, like most Indians, though I started programming at the age of 12, was lucky enough to have a ah computer back then.
The Evolution of AI
00:01:45
Speaker
fell in love with programming, did a lot of it through school and college. As you said, I went to the US and did a master's in AI. This was back in 2009 before, you know, the current AI explosion as we're seeing it.
00:01:57
Speaker
i mean, this was the era where things were, expert systems was the way in which people thought of AI.
The Rise and Fall of Chatbots
00:02:03
Speaker
Yeah. Expert systems, but also machine learning. Machine learning was um quite hot as a space and it was being used for very siloed applications. Like you'd see machine learning systems for um personalization algorithms, for example, on Netflix or e-commerce websites.
Pivot to Job Placements
00:02:21
Speaker
right That was a big, big application. ah you you were seeing hints of conversational AI kind of a appear.
00:02:31
Speaker
In fact, there was a big chatbot boom back in 2016, 2017 when we actually started Bahan.
Inspiration Behind Vahan
00:02:37
Speaker
And then it died down very quickly because the AI and the tech just wasn't ready.
00:02:43
Speaker
ah you know and It's come a full circle even for us, which I'll talk about. So yeah, so going back to my journey, studied in the US, studied AI, ah post that I worked at various startups in the US for about seven years, across the board, you know, 500 people ah post product market fit and also five people pre product market fit. So I learned a lot about this building companies in general.
00:03:07
Speaker
Were these consumer tech businesses, B2B SaaS? What kind of businesses? A couple of them were in the consumer space. There travel company called Jet Setter in the consumer space.
00:03:19
Speaker
ah which actually did really well, got acquired by TripAdvisor. And then there were a couple in the B2B space as well. um And I remember working at these companies, this all in the US. And one of the great things I think about just culture over there is that everybody's allowed to be ambitious and, you know, do as much as they they want to. You're not really slotted into clear, you know, limited roles.
00:03:44
Speaker
I always knew that I wanted to start a company at some point. So I took it upon myself to learn everything I could, not just about building products, but also running sales, running marketing, building culture, hiring people, all of those things.
00:03:58
Speaker
Very, very enriching, very rewarding experience. In 2014, I had this incident in my life that really caused me to reflect deeply on what I was doing.
00:04:10
Speaker
ah I was visiting my parents in Delhi at that time. And one of the things we did fairly often ah was distributing food to people who just passed by outside our house. And we live at ah at a very busy intersection in Delhi.
00:04:25
Speaker
um we usually cook something like you know khichdi or dal chawal and we were handing this food out and I remember giving a plate of food to a kid who must have been 11 or 12 years old tackered clothes and he took the plate of food from me and gave me you a big hug in return And I remember even now when I think of that, ah you know, it sort of makes me very emotional. Something just snapped inside me.
00:04:50
Speaker
and was literally in tears in that moment ah because I realized how fortunate I was to be handing the food out and not receiving it. um And in fact, there's a very startling statistic around ah around this. It turns out that if you have a roof over your head, you have access to the internet, you have a college degree.
00:05:10
Speaker
And if you can afford three meals a day, these four things, then you're part of just 7% of the world's population. oh So where you were born, if you think about it, is like a lottery ticket.
00:05:20
Speaker
yeah Warren Buffett, in fact, he calls it the Ovarian lottery. He's he's coined that term. Most of us on this, so you know, who are listening to this are probably winners of that. um So we started Vahan, or at least the inspiration to start Vahan came in that moment.
00:05:36
Speaker
I wanted to use technology to essentially break the Ovarian lottery in that moment. At this time... You were working at Adstruck in the US. was Adstruck?
00:05:49
Speaker
Adstruck was a B2B platform for outdoor advertising. So we built a software that would make it easier for advertisers, the agencies or brands to advertise using out of form. So billboards, essentially, was the first of its count.
00:06:07
Speaker
This was like the software part of it was what will be displayed or the bidding... part of it and like Software was ah more um finding the right ah media. Okay, right. right The buying weed buying process. Correct, the buying part of it. We'd aggregated the supply. I think we had over 85% of the supply on our platform.
00:06:28
Speaker
And then we would make it easier for advertisers and buyers to find the right media for their budgets. Okay, got it. Okay. So, ah like did you raise funds, etc., being in the US but before starting? Or like what was the way in which you decided you want to do something that impacts the poverty you saw in India? So, what next?
00:06:53
Speaker
So um one of my friends who lived in the valley at that time, I got talking to him about this problem statement. And he was, in fact, ah volunteering for an NGO called Hippocampus Learning Centers in Bangalore, which does some work around training rural women to teach maths and English within their communities.
00:07:16
Speaker
And we started talking to them just to understand what types of problems they were facing. And we realized that teaching these women English was a big problem. And the reason that that existed was that these women didn't really have a way to practice their English. And this is true for all language learners.
00:07:36
Speaker
but If you're using an app to learn in any sort of language, you'll typically be limited to you know practicing that language with that app. But you're living in a community of people generally who speak you know your native tongue.
00:07:49
Speaker
And so you never get to practice. So in fact, the first product that we built, and this was at that point a side project for me, we built a product called Lakshmi. which was a virtual English teacher whom you could call on the phone. You could chat with her.
00:08:05
Speaker
And she would ask basic questions like, what is your name? Where are you from? What do you do? And we had very basic AI technology at that time, speech recognition that would try to recognize what the person has said and give them feedback.
00:08:18
Speaker
So it was a voice bot, right? Back in the day. Okay. And how this would operate, I imagine, is you would transcribe the response, like whatever the person is saying would get transcribed and ah then there would be some sort of ah accuracy check and then feedback given on whether it's... Exactly.
00:08:40
Speaker
So it would be unable to fix things like, say, a mother tongue influence that accented related issues and stuff like that, but basic grammar... And I'm sure even transcription would be like 80% accurate at best. Exactly. Exactly. Exactly. These systems, again, were, I think, a far cry from what we have ah but available today.
00:09:03
Speaker
But the idea was essentially to give people a natural conversational experience or at least a way to... just stood you know practice as much as they can and you know later when we deployed this at vocational training institutes in India ah there were people who would use this for hours and hours any given day I remember the story of a girl I think her name was Santoshi based ah in Karnataka who sat with this so and, you know, I think she would ask you maybe 10 questions, the English teacher.
00:09:39
Speaker
She went on loop, you know, with these 10 questions for like six hours. Hmm. And when we spoke to her to understand why she said, you know, I don't even get a chance to go outside my home and like interact with people. So this is a way for me to just interact with someone, you know, obviously practice English.
00:09:58
Speaker
ah But it also became an outlet ah for some of these people.
00:10:05
Speaker
So why aren't you running that today? What happened? yeah So we've gone through an interesting journey through the lifetime of Ahran. So I ended up moving back to India when when I built Lakshmi.
00:10:18
Speaker
We tried deploying it in the US at a few ah community-based organizations that were teaching English to immigrants. And ah in parallel, we deployed it at Hippocampus in Bangalore.
00:10:33
Speaker
Eventually, i realized that if I really wanted to make a dent, I would have to do this full time. ah So i gave up the life, ah you know, the glamorous life of New York or whatever, came back to India, almost straight to Bangalore and started deploying this product in vocational training institutes.
00:10:52
Speaker
And if you remember back in 2015, 2016, skilling was getting a lot of attention. ah The National Skills Development Corporation was formed around then. The PMKVY, you know, Tadhan Mantri's skilling ah program, etc. took off.
00:11:10
Speaker
And India does have a massive skills problem. I believe we produce seven to eight million graduates a year, but more than half are unemployable. And a large part of that gap comes from their inability to communicate in English.
00:11:24
Speaker
Yeah. So we said, okay, let's help people learn English by practicing and let's tie up with vocational training programs to give them Lakshmi as an aid for their students.
00:11:36
Speaker
So we started there. We got a few paid pilots. um We got really good engagement on the product. People were using it. It was working well. But then those paid pilots didn't really convert into recurring contracts, long-term recurring contracts.
00:11:53
Speaker
And we realized that skilling institutes weren't really incentivized to deliver better learning outcomes. They were really incentivized just to place the students that they were training, quote unquote.
00:12:08
Speaker
um And so from there, we decided to pivot to a sort a slightly different market. ah We got introduced to Uber, who told us that they wanted to train their drivers and not just in English, ah but in other important skills, such as ah using Google Maps or using the Uber app better or how to treat their customers better.
00:12:32
Speaker
So we started building at that point. And by the way, Uber said, Hey, can you train 10,000 on my drivers and I'll pay you a hundred rupees a driver. So it was a 10 lakh rupee pilot. This was just a pilot.
00:12:43
Speaker
We pivoted overnight. and We said, great, there's money. Let's pivot. it So far, this built out of your savings. Yeah, this was no funding raised ah so far.
00:12:55
Speaker
And then what happened is... And which year is this when the Uber pilot... i think this is this is about 2017. We had at ah at some point during this journey, we had received a small grant from an organization called Vilgo.
00:13:11
Speaker
Very, very small grant. um these ah So Vilgo funds of social enterprises. Okay. And they give very small seed, pre-seed grants. So it was that money and basically my savings.
00:13:25
Speaker
ah We pivoted um to this bot for training ah corporate workforces, right? Like Uber drivers. Yeah. And we also realized that WhatsApp and smartphone penetration was taking off in India.
00:13:39
Speaker
So we integrated the bot into WhatsApp. And this, by the way, was pre-WhatsApp API. Yeah. And there's still voice ah because like if you want to teach someone how to use Google Maps, how do you do that through voice?
00:13:51
Speaker
Yeah. So the WhatsApp integration was text and it was multimodal essentially. Okay. um We actually botted the WhatsApp API. ah Actually, there was no API. So we botted the app itself.
00:14:06
Speaker
What does that mean? Botted app? So we built automation that would be programmable. Not really jailbreaking. when For example, when we wanted to send a message through WhatsApp, and we would write a ah snippet of code.
00:14:23
Speaker
And that code would call some other code that would ah open up a WhatsApp instance on an Android emulator, type in the message, click the button and send it. Okay.
00:14:34
Speaker
And this would be repeated thousands of times. Correct. like like go Okay. It was a workaround because WhatsApp was not selling API. There was no API back then. There was no API back then. So we built this this botting system pretty much for all WhatsApp functionality for sending audio messages, for images, text, obviously receiving messages.
00:14:57
Speaker
i think we were probably one of the first companies to do this kind of work. Um, And so we integrated Lakshmi into WhatsApp at that time. And we started building verticalized courses, for example, for Uber, for their drivers.
00:15:12
Speaker
We did a lot of work with Flipkart for their delivery workers. We did a lot of work for Club Mahindra for their salespeople. And this was like English and ah functional skills plus English.
00:15:26
Speaker
Functional plus English, exactly. Job-specific functional skills plus English, verticalized courses. So we had some elements of product market fit. We were getting revenues. um What kind of revenues? Like 2017, that financial year, how much would you have done?
00:15:43
Speaker
Oh, it's a long time back. i I really don't remember the exact number, but let's say off the order of maybe 50 lakh for the year or something of that sort. Right. So it was definitely much better than what we had before.
00:15:55
Speaker
Yeah. And we were starting to see early signs of PMF. What we realized eventually that um even in this space, we weren't getting...
00:16:07
Speaker
long-term recurring contracts. While customers were willing to sign up with us for six months or maybe a year at best, they weren't willing to give us a five-year recurring contract, you know, with an annual subscription fee or something of that sort.
00:16:21
Speaker
And when we tried peeling the onion there, we realized that training ultimately is really nice to have. It's ah a vitamin in a way. Whereas the real painkiller for these companies was to hire people right because they were hiring thousands of people every month.
00:16:39
Speaker
And typically in the blue collar space, churn is also very high, which makes hiring an even bigger pain point. And when you look at the blue collar worker side or the user side, the dynamic is very similar.
00:16:52
Speaker
Learning something is, again, a nice to have. For them, getting a job, having an income, being able to put food on the table is exactly, is the painkiller, it is the real painkiller.
00:17:05
Speaker
And um we did a pilot back then with the same tech stack with the company called Dunzo, which was ah you know also pathbreaking company based out of Bangalore. I think they pioneered the entire Q-commerce model also whatsapp based or at least in the for a large part in the early days they were yeah ah they were ah in indramagar as well where our office was and i remember chatting with one of the founders and he said hey like we have all these leads of delivery workers whom you know we call to hire and can you do something with your whatsapp bot to filter them out for us you know to reduce our hiring costs
00:17:48
Speaker
So we ran a quick pilot with them and we were able to reduce their hiring costs by 30%. What was that pilot? Like a voice bot who's interviewing these delivery workers? It was a very, very simple text-based bot at that point okay where we would, yeah, we would interact with the these candidates on what WhatsApp text, ah get basic details like where whether they were based, what their name was, et cetera. And then whether they were interested in the job or not, right? Give them details about the Dunzo job and then whether they were interested.
00:18:20
Speaker
And then we would pass... interested leads, essentially a set of qualified leads yeah over to the Dunzo team, right? To to convert. And so that, just that activity was able to reduce hiring costs by 30%.
00:18:34
Speaker
So we felt that we were onto to something and suddenly we started to get pulled from the market. Danzo referred us to Swiggy. Swiggy referred us to someone else and then to someone else. And ah you know Mark Andreessen actually says this, that PMF occurs when you start to get pulled from the market, when customers just start referring you to each other right and you start to get inbound.
00:18:56
Speaker
And we saw that happening. And so I would say that's really when we sort of hit PMF, when we were solving for a deep rooted pain point for a you know specific segment of the market.
00:19:08
Speaker
When did the Denzel pilot happen? Which year? This ah would have been around 2018, around that time. Okay. And yeah, go ahead. couple of questions on this three year journey of finding PMF. ah Yeah. How were you opening doors? I mean, you know you worked in the US, you you yeah not really, ah ah would not probably have seen yeah how sales to enterprises happens in India and stuff like that. yeah Like how were you, yeah ah like right from the skilling institutes which you were selling to Dunzo and all, like how was that happening? yeah yeah
00:19:50
Speaker
Yeah, that's a great question. um it was It was a learning curve for me for sure. I was doing all the sales, obviously, as a founder and you know solo founder.
00:20:02
Speaker
ah Most of the introductions happened through the network. um And I realized actually that the network in India is fairly tight. It's not very, very difficult to get introductions to almost anybody you would want to through the network.
00:20:20
Speaker
Even the scaling skilling organizations? I'm sure the startup network is tight. Like you can ah like leverage one founder to connect to other yeah founders.
00:20:31
Speaker
Yeah. I think I got introductions to the skilling organizations through ah the social enterprise community in Bangalore. So Bangalore then had a very, very tight...
00:20:43
Speaker
social enterprise community. um A lot of companies who were, let's in the financial services space or livelihood space. ah Through those people, I think I was able to get the right introductions.
00:20:55
Speaker
okay okay We also had the backing of WillGrow. So I think they helped ah open a few doors for us. ah But then I remember...
00:21:07
Speaker
still relying on back then selling value. ah You know, the I think this is maybe a sort of and a newbie or a new sort of mistake that one one makes when one is doing sales. There's a lot more to sales, especially enterprise sales and just selling value, right? While it perhaps starts there.
00:21:30
Speaker
There's a a lot of ah dynamic around understanding the incentives of the person you're selling to, of who influences that person, of understanding that the buyer and the decision maker or the user and the decision maker might be different people. Sales in India is also very, very top down. That's been one of my biggest realizations.
00:21:56
Speaker
In fact, one of my mentors tells me that if you ever want ah you know your your customer to do something for you, by customer I mean the actual buyer, that person should know that you know their boss.
00:22:10
Speaker
Yeah. It's very, very top down in India. And ah that's a ah much faster and more efficient approach to sales in this country.
00:22:22
Speaker
What do you mean by ah that newbie mistake of selling value versus what you do today? What's the difference? Yeah. Selling value would be like, i can help you to reduce your hiring costs, right? That's clearly like the value.
00:22:39
Speaker
So, so what actually yeah yeah, help me understand the difference. here Yeah, so exactly. So if I were to sell the recruitment product, right, I would say, hey, here's a product. I can reduce your hiring costs by x percent, which will save you, you know, Y in rupees or dollars or whatever it is, right? Great. That's value, right? ah But does that person even care about that?
00:23:01
Speaker
okay um right That person ah may have a mandate to just, I don't know, grow their top line and not reduce costs.
00:23:11
Speaker
Or that person might have a mandate to, I don't know, expand into certain geographies. um And in fact, what I learned was, broadly speaking, that generally in India,
00:23:25
Speaker
cutting costs is, especially when it comes to labor, ah is not a very strong value problem because cost of labor is is low in India. And a lot of SaaS software therefore fails in India because SaaS typically results in human labor reduction.
00:23:45
Speaker
which works in developed countries, you know, obviously very well, um especially in the U S because the U S in general is a very, very productivity and efficiency focused economy and, and culture.
00:23:58
Speaker
India is not like that, you know, so that would be selling value. But tying that back to what that person cares about and also tying it back to perhaps what their boss cares about. I found that to be ah perhaps slightly more important in a country like India.
00:24:18
Speaker
how do you how do you discover what you need to sell because what you're saying is you cannot pitch your ah offering without first understanding what is what the customer cares about so ah how does it start then Yeah, so ah what I learned is that in sales, especially when you're initiating conversations, it's a lot more important to ah listen and let the customer do most of the talking.
00:24:53
Speaker
And I think, again, this is, ah I think, a new mistake. I think a lot of us, when we get into sales, we just try to sell. like Just take whatever I have. It's that that classic sell me this pen sort of scene from the Wolf on Wall Street, right? I might say, hey, I might come to you and say, this pen is awesome.
00:25:12
Speaker
You know, it'll it'll help you write the most beautiful a sort of handwriting and, you know, it's easy to write with and it'll never run out. But I didn't ask you, are you even in the market for a pen?
00:25:25
Speaker
If you're not in the market for a pen, then like I'm wasting my time. Right. Right. And so it's so important to let the customer do the talking and understand their world much better. And when I say their world, it's not just that one individual, but influencers and decision makers around those people, right? That that mini ecosystem, so to speak.
00:25:49
Speaker
um and like political dynamics and you know who really takes the calls and uh uh you know who influences who and how to build champions within this mini ecosystem and things like that um right so i think that is is also very very important aspect of this how do you initially get them to give you time like you would need to pitch something first right to to get time from them how does that happen So the the best way to get time is through a warm introduction.
00:26:19
Speaker
yeah Though I remember doing a lot of cold emails, a lot of cold calls and just trying to establish credibility. um That's also one thing that I learned about sales early on that any sales pitch, especially if it's a cold pitch, but even if it's a warm one, ah it needs to start by establishing credibility.
00:26:40
Speaker
You need to talk about the brands you're working with, like for credibility. That could help, right? So credibility can be established through um social proof. Social proof could be, hey, I already work with these brands. Or it could be, hey, I know so and so who's referred me to you. Or ah it could be your own background. So in my case, my background also helped because I was US return, quote unquote, right? I studied Columbia, etc.
00:27:06
Speaker
Good school, good background. I think my communication always helped because I speak relatively decent English. i So then i building credibility is also very, very soft art, I would say. sort of And so once you establish credibility, you're able to then ah it least get the person's ear.
00:27:25
Speaker
And that would generally help me at least land a pilot in a lot of cases, or at least help me get the customer to talk, you know, talk about their challenges, their issues, ah what kind of product would really help them.
00:27:42
Speaker
Okay. Okay. Let's continue with your journey. So 2017, Dunzo unlocked a new market for you. You found PMS. So 2016 was working with the vocational training institutes. 2017 was working with the likes of Uber, Club Mahindra doing skilling, etc. 2018 was pivoting to the job space.
00:28:05
Speaker
Okay. Let's see the likes of Dunzo. And even within the job space, I think we we had like many smaller pivots. We realized that processing leads, which is what we did for Dunzo, we qualified leads, right?
00:28:19
Speaker
Was, ah again, less of a value add than actually generating those leads for employers. Right? So Dunzo would be willing to pay us, I don't know, back then, five rupees or 10 rupees per lead processed.
00:28:32
Speaker
which was a commodity business, right? We would not make a lot of money, but if we actually brought in new leads, they would pay us a lot more. So we started trying to crack that problem. How do we generate leads on our platform versus, you know, just qualifying leads.
00:28:48
Speaker
And we started initially doing a lot of performance marketing on the likes of Facebook, to bring in job seekers into our funnel. ah How were you funding the Facebook ad? like electronic at that point ah yeah At that point, we'd raised a little bit of angel money.
00:29:06
Speaker
um Gokul Rajaram was one of our first, in fact, he was the first angel investor in our company. Gokul is perhaps the most prolific angel investor in the world or one of the most prolific angel investors.
00:29:18
Speaker
He was the guy who built AdSense at Google. So that's, you know, one of his, um i think, bigger achievements. And then he's worked at DoorDash, a lot of great companies, Meta as well.
00:29:32
Speaker
ah So Gokul's a great guy. He put in a small check and then we had a few others. Again, you know, Gokul helped us establish credibility. And then, you know, we sort of raised a little more angel money back then.
00:29:44
Speaker
Mostly US, like the investors. Yeah. Some US, s some in India as well. Okay. ah Back then, I think Veer Kashyap, who was the founder of Baba Job, which is also a large blue collar platform, he put in some money.
00:29:58
Speaker
We got, I think, Mekin Maheshwari to put in a little bit of money. Mekin was the head of engineering at Flipkart back then. Flipkart, you know, was doing quite well.
00:30:10
Speaker
um So it was a mix. And so we started spending some of that money on ads. But we realized it was very, very expensive. Very, very expensive. I can imagine. this Yeah.
00:30:22
Speaker
Yeah. And the the economics were just not working out. So we tried to figure out ways through which we could generate more organic interest on our bot. And at that point, we had a unified bot called Job Finder that anybody could message and it would match this person with jobs and, you know, get them to apply.
00:30:43
Speaker
um This was a WhatsApp bot, this job. This was a WhatsApp bot. This is also pre-WhatsApp API. So WhatsApp API came out, I think in February 2019. That's when we got access to the official API. So this is all through that bot automation, you know, that I spoke about.
00:31:00
Speaker
So um we started experimenting with a referral program. You know, we said, hey, can we get people to just refer other job seekers onto the product? And ah we were doing that, trying various experiments.
00:31:14
Speaker
um We would send out a message right after you would apply for a job on WhatsApp that would be easy to forward to friends. And anybody could kind of click on a link in that message and then chat with the bot.
00:31:29
Speaker
And if they applied for a job, then the person who referred them would get some money. You know, it was an amount like five rupees or ten rupees. And I remember this even now quite well. i think this was in, ah this was a around late 2018, early 19. Our referral program went viral.
00:31:51
Speaker
There was one day where suddenly we started to get a massive influx of users. We were getting, i don't know, over a lakh every 10 minutes.
00:32:03
Speaker
Wow. Yeah, it just blew up suddenly. And we just could not handle the traffic. We tried scaling our servers, doing a bunch of things. But eventually we just took down the bot and we just put a form out there so that we wouldn't lose people's details because the traffic was just crazy.
00:32:20
Speaker
And this thing just blew up. It just suddenly went viral. How are you doing these payouts? Like that sounds like a painful thing, like like that microtransaction. So, Paytm. Paytm had already become quite big back then. okay So, we were doing these payouts to Paytm, connecting to people's Paytm wallets.
00:32:40
Speaker
which worked which worked quite well. And oh this is pre-UPI, right? um And then at I think at peak, we were generating at least half a million to a million MAUs.
00:32:53
Speaker
Again, this was probably February, March 2019. And that's when we applied to y Combinator. And an MAU, how do you define that? Like somebody who applies for a job or like what what's it?
00:33:08
Speaker
Yeah, somebody who is essentially coming to our product and applying for a job. And the product is a bot? The product is a bot. on On WhatsApp. So they would just chat with it.
00:33:18
Speaker
It would gather some basic information and recommend jobs. Okay. Okay. Got it. Okay. Yeah? Yeah. So then now we this was early 2019. We applied ah to YC and we had decent traction. We were generating revenues. We had traffic.
00:33:37
Speaker
The referral program was blowing up. And I think they had their interviews for the first time in India that year in Misho's office.
00:33:49
Speaker
10 minute rapid fire interview two partners it was quite ah a world whirlwind experience and I remember getting a call that night from the US saying that hey you know we'd love to fund you and that was just a surreal moment I mean you getting into YC is for any entrepreneur almost like you know getting into Stanford or Harvard it's probably better than that in a sense in terms of the you know <unk> excitement And, of course, accepted, attended YC summer 2019, which was, again, a life-changing experience. um As in, you moved to the US to attend? it No, no. So they didn't. I think there were 12 companies from India in that batch.
00:34:35
Speaker
Nobody had to move to the US because it was, you know, your business was in India. So we would all keep sort of traveling back and forth. I'd go there, spend two weeks at a time, come back, spend time with the team, then go back, um you know, so that I could interact. You'd go there, you'd network, learn some something, get some insights, and then come back and implement. and Exactly.
00:34:58
Speaker
Yeah. Yeah. And I think the biggest value that we got from YC, other than the, you know, the brand association and recognition was um the network. The YC network is incredibly, incredibly um smart, very capable and also powerful in Silicon Valley.
00:35:20
Speaker
um So getting access to that that peer group, being a part of that peer group um was just amazing. And YC has also built this culture, which is very special to YC as an incubator, ah where people in the community really do go out of their way to help each other.
00:35:39
Speaker
ah And you'll find that you know even if you write to people like Brian Chesky or even Sam Ortman, they'd respond. um So Sam, in fact, ah even now when I write to him on email, he'd respond, might take some time.
00:35:55
Speaker
But I always get you know a one-liner back or something. ah So the the the network is just very, very dense and incredibly powerful and strong in the very...
00:36:07
Speaker
So YC happened in summer of 2019. And then I would say we raised our first ah significant round of money. ah Mostly angel investors putting in smaller checks.
00:36:21
Speaker
um And Khosla Ventures, in fact, put a small check as well in that post YC round. That was our seed round. um yeah and then it was sort of off to the races how much value were you capturing at this stage so you told me like for qualifying leads you would get 5 10 rupees per qualified lead ah for sourcing ah like like what kind of ah pricing was it or I think for sourcing we were getting maybe about 50 or 100 rupees or something of that sort ah you know
00:36:52
Speaker
And in fact, ah that's a good point because we, at that point, we have to make another business model pivot. it We realized that even selling like leads and qualified leads wasn't capturing a lot of value.
00:37:07
Speaker
And the real way to capture value with the enterprise segment in India was to place people. So enterprises were more than happy to pay us. Correct. We're more than happy to pay us thousands of rupees if we actually got people hired.
00:37:22
Speaker
um And we were trying to push this lead model for a long time because that's what ah what has worked in the West, right? Like if you look at companies like Indeed or Monster, um is typically in the the white or gray collar space, right? They sell resume resumes or just qualified leads or applicants. They don't sell hiring, right? Which is to the point, the agency model.
00:37:45
Speaker
But in India, especially in the blue collar space, the agency model is really what works. ah employers don't really pay much for anything else. And so, ah you know, we we debated internally a lot whether or not to switch to that model because the data on who's actually been hired would come from the employer. and There would be challenges around closing that loop.
00:38:08
Speaker
But we bit the bullet. We said, hey, like, if this is our tool this has to work in India, it's the only way it will work. And so we pivoted to that model. I think at that point, Zomato was one of our first customers under this model where they set up some sort of automated report to get us that that employment data on a daily basis.
00:38:29
Speaker
okay So we started scaling this model at at that point. Which year was this? This was so early, 2020. Post YC. This was post YC, yeah. this is voice i see I'm curious about something. Yeah.
00:38:45
Speaker
So there is this Apna, which is the only unicorn in this hiring space, ah which is essentially selling leads. Yes.
00:38:57
Speaker
What is the reason why you managed to... ah like Like move up the value chain and not think of following up NASPATH because there there is clearly a path to becoming a unicorn just by selling leads.
00:39:14
Speaker
Yeah. what what What made that happen? what Was it the fact that ah you were pretty much bootstrapped till that point of time and so there was no investor pressure to...
00:39:26
Speaker
chase a valuation or what was it like? Like just fundamentally, structurally, what was different that you did not end up following? so yeah. So fundamentally, we've never been a company that chases valuations. We've always been a company that wants to build a sustainable, large-scale business that actually generates, you know, margins.
00:39:50
Speaker
There are so many Deeper thing here, you know, I'm sure if I interview the upna founders, they will give me the same answer that we are not a company which changes valuation. Nobody says that. So ah like what?
00:40:05
Speaker
So, I mean, so Apna also, by the way, the valuation that they got wasn't on the basis of revenue. So when they got that, you know, billion dollar valuation, this was 21. Yeah, it was on engagement, ah just pure retention on their product. And retention was driven by a community feature on their product.
00:40:25
Speaker
um they were pre-revenue and even today ah they don't really have significant revenues to talk about because they haven't been able to figure out that business model that will really scale in the space in India and capture enough value where you can generate profits so that's what I'm trying to figure out but why did you figure that out what What was different? We could have, we could have ah by the way, raised on the basis of valuation, basis or on MAUs, right? Because um using our referrals were blow blowing up. We had a lot of usage on our bot. But ultimately, you know even then, we we weren't finding a lot of placements happen through the bot.
00:41:06
Speaker
Even though we had millions of people coming in and applying, a lot of people weren't getting placed. and And a very core insight that we uncovered around that time, Akshay, was that blue collar workers in India don't look for jobs online.
00:41:23
Speaker
it's still a very, very human-driven process. And um there's a larger market level insight here about India in general. I found that in use cases where there's a lot of complexity or there's a lot of money to be exchanged, um there's typically...
00:41:44
Speaker
an offline behavior where an agent or a middleman might be dominant. um So car sales, for example, right, in India are still largely done offline insurance. Yeah. offensive right ah Equity buying, by the way, is still done largely through brokers because it's complex.
00:42:02
Speaker
Travel bookings, by the way, are still done largely through travel agents in India, even though you have OPAs. um In fact, insurance sales, financial services, at all these are done largely offline.
00:42:19
Speaker
ah So jobs is very similar in the blue collar space. And most blue collar workers will find jobs either through their networks, through their mama, al ja ba yeaha their friends and family, or through local agencies.
00:42:32
Speaker
Because it's it's a complex transaction um and it requires a lot of trust. ah yeah The other example I had was ah e-commerce. but If you remember, back in the day, e-commerce was not taking off ah till the COD, the cash on delivery model was introduced.
00:42:50
Speaker
yeah yeah Flipkart had to introduce the COD model to bootstrap the space and to build a level of trust and credibility before people started ah paying with credit cards.
00:43:03
Speaker
And even now, by the way, pretty large fraction of these transactions are done through cash-on-delete. Okay. Fascinating. So essentially like India is a low trust market. So there's a crossmark you need an agent of trust to, you need a human being at some point in this transaction to be able to bridge that trust gap and consummate the transaction.
00:43:27
Speaker
And um which is why, by the way, we had a ton of traffic on our bot, but very few people actually getting jobs. Hmm. ah ah Was it partly your, I mean, the reason for you to be an entrepreneur was to move the needle for this segment of society and you saw that the needle was not moving and that made you want to pivot your business model?
00:43:54
Speaker
yeah Ultimately, our goal was to get people jobs through this product. right And that's how we would capture value because employers, our customers would pay us for people who were actually getting placed.
00:44:05
Speaker
And it was obviously, I would say, part mission link, but our business model was also tied to that impact as well. you know ah So the market pulled us in that direction too.
00:44:18
Speaker
um And if people people were applying for jobs, but they were not getting placed, one, we weren't creating any sort of impact on society, but we also weren't getting paid. What was the ah changes needed? you You would have to retool your business to, yeah I mean, agency model, it's called agency model because it's people running it, right? yeah the Otherwise, like like it's not like software can run it. So you would have needed but to retool your organization and how you work to ah to you know align with this new business model of pay per placement instead of pay per lead.
00:44:58
Speaker
So the problem statement at that point obviously was how do we get enough job seekers into the platform in a way where they actually end up trusting the platform and they actually end up getting placed.
00:45:15
Speaker
And the key insight, like I mentioned, was that this process is still very human driven for people. So we said, can we build a product for that middleman? Can we build a product for that person who bridges this trust gap?
00:45:27
Speaker
And we started experimenting with a bunch of things. um Eventually, what we ended up doing was building a product for local recruitment agencies.
00:45:38
Speaker
And it's a lesser known fact, but in India has hundreds of thousands of local recruitment agencies who do close to 60 to 70 percent of all blue collar hiring in this country.
00:45:49
Speaker
And these are typically solo entrepreneurs run with a team of maybe five to ten recruiters. And these recruiters will be on the phone all day or they might be in the field trying to recruit people.
00:46:01
Speaker
So what Vahan has become over time is essentially a marketplace where demand comes to us in the form of jobs from large employers and supply actually comes from and pop type recruitment agencies.
00:46:16
Speaker
So they use our software to discover demand, job opportunities, to match candidates with these job opportunities. And the AI that we're building, in fact, the the latest product that we built is a voice bot that talks to candidates on the phone in Hindi and English that qualifies them, that answers basic questions, who did it collect documents and things like that.
00:46:41
Speaker
um This is actually a tool used by agencies. to increase their efficiency and productivity. and So most of the software now that that we've ended up building actually helps agencies do a better job at hiring.
00:46:58
Speaker
You had ah half a million MAUs at one point a time. ah What you're describing to me now is PR, B2B, both sides, your supply is coming from...
00:47:11
Speaker
ah small businesses which is these agencies and your yeah demand is coming from large enterprises who want to hire so yeah i mean as a founder i would feel like ah i have this meu i should leverage that i should sweat that asset ah like yeah what happened to that So one, by the way, another interesting realization for me was that um engagement. So MAU is essentially engagement, right? Or or attention.
00:47:45
Speaker
In a country like India, that doesn't always translate into a business. the The kind of default playbook in the in the Valley is, hey, you you build engagement and then that will translate into revenue.
00:47:59
Speaker
ah That usually works in developed countries develop because people have more money to spend, more discretionary spend, etc. right So that playbook works. That doesn't work in India. um WhatsApp is sort of the poster child for this.
00:48:12
Speaker
So WhatsApp is perhaps the most used app in India. I think some 80% of Indian smartphone users use it every single day. ah But when WhatsApp tried to charge people literally 50 rupees a year, nobody paid.
00:48:28
Speaker
And which is why WhatsApp ended up building a B2B business model. Because that was the only way to monetize the platform. There are many such examples, even if you look at the likes of Facebook, Meta, etc. Instagram, right?
00:48:43
Speaker
India is like a DAU, MAU farm for them. ah Their ARPU in India is like one-tenth, if I remember correctly, one-hundredth of what it is in the U.S. ah Because in India, people in India are time rich, but money poor.
00:49:00
Speaker
and So they give you engagement, they give you time, but they may not, as consumers, give you money. like So I realized this at that time that, look, look we can build a ah large sovereignty metric in my mind. right You can feel good about it you know yeah but it won't really lead to a business.
00:49:18
Speaker
So you stopped that completely, like ah in terms of trying to get referrals and You could juice your margin by taking a hybrid yeah juice your margin by taking a hybrid approach like So like one approach is that all sourcing will be handled through my agency partners. They will shortlist and give me candidates and I will present those candidates to the recruiters and take a margin from that and pass on some money to the sourcing partner. Or the second is that maybe some of my own MEU traffic
00:49:56
Speaker
will also get hired and I don't have to pass on anything to an agency partner for those recruitments. Yeah, yeah, yeah. Yeah. So um in fact, we are now sort of building a hybrid to your point.
00:50:12
Speaker
ah we We now have acquired over i think, 30 to 40 million candidates through the agency approach. And we're now, in a way, trying to engage that base of candidates to see if we can juice, to your point, right those people.
00:50:33
Speaker
um So I think it was ah it was also a very conscious call to go about this in a very staged manner. where we said the agency model is a much, much better way to not just get distribution, but to also build a business and a viable business, by the way.
00:50:51
Speaker
ah The economics of this business also work out well. um And then later, once we have acquired a lot of users ah through this channel, we can re-engage those users and do some of this directly.
00:51:06
Speaker
And I always feel that in a country like India, given the size, the scale and the social dynamics, both these models will coexist. Okay, I guess ah maybe the technology also was not good enough for you to be as effective as an agency till now.
00:51:24
Speaker
That's also ah ah a valid point, right? It's come a full circle for us where our first product, if you remember, was a boy voice bot for English, but it was a voice bot. And we've ended up building a voice bot now as well for recruitment because the tech is just so much better.
00:51:39
Speaker
ah You're now able to deliver a much more humor-like experience than you could back then. And that is critical for this product to resonate and to work with this audience. What is the... ah Like, you know, from a user experience perspective,
00:51:59
Speaker
what would a close to human experience need you to build? ah Like, are we talking of instant reply that as soon as ah an answer comes, the next question is immediate and therefore you need on the fly transcription processing, responding, like like that latency, low latency or what yeah else is there? like So latency latency is definitely a part of it.
00:52:22
Speaker
I think ah broadly I would term the it as just empathy. um right um So India is a very diverse country. You know this, right? 22 major languages, thousands of dialects. I mean, every hundred kilometers, how people, you know, dress, speak, eat, just completely changes.
00:52:42
Speaker
And delivering a human-like experience then at some level means being able to capture that diversity and being able to empathize with the person you're speaking when it comes to their cultural context and background, right? So language becomes important more than language, dialect and accent become important.
00:52:59
Speaker
Cultural references might become important. ah And that's how you establish trust ultimately as well. right So those are the nuances that we are trying to design our product around.
00:53:10
Speaker
um you know I was talking to my team about designing this bot. And I gave them a very simple example. I said, hey, when you meet a stranger, usually in India, you end up asking where they're from.
00:53:24
Speaker
That's a very, very common question. Where are you? And then they might say, oh, I'm from so-and-so hometown or you know place. And you'd be like, oh, you know, my cousin or my mama or somebody I know is from there or maybe my know parents were from there.
00:53:38
Speaker
Yeah, that same thing, which in sales, you told me ah that credibility building through social And suddenly you don't know each other, but suddenly, you you know, you you'll bond.
00:53:49
Speaker
ah Because, you know, some distant relative was from the same place. And I was like, you know, the bot needs to do that. The bot needs to, these are the ways through which Indians find common ground and ultimately, you know, create some amount of trust and empathy ah with each other.
00:54:07
Speaker
Mm-hmm. So ah how the let's go a little deeper into the agency model with offline agencies as your sourcing partners. How does it work? ah Just take me through the whole workflow.
00:54:21
Speaker
Yeah, so ah these agencies find candidates ah for jobs through multiple different mediums. They might find them through their networks.
00:54:33
Speaker
ah Some of them run um ads at local you know pawn shops, chai shops, stuff like that. Some of them will buy leads ah online um and then they have a team of recruiters. ah They'll either be on the phone um calling these people and trying to pitch them different job opportunities and then eventually you know help them get onboarded.
00:54:53
Speaker
Or there might be people in the field who will actually handhold these people, maybe take them to the employer's location, get them onboarded and stuff like that. um So they they use our software to one, discover what job opportunities are available, um which ones are relevant to them.
00:55:11
Speaker
ah They'll match candidates to these job opportunities based on you know candidates, information, background, et cetera. Like the phone number of the candidate would be the unique ah identifier. is Typically the unique identifier. Exactly.
00:55:26
Speaker
They'll be able to track where the candidate is in the hiring process in near real time. So we've built now integrations with most of our customers where this data comes in and it's reflected on our software.
00:55:39
Speaker
So then the recruiter knows at any given moment in time where X candidate is and what they need to do with that candidate. Which is also, by the way, i think a very unique capability that that we built. I haven't seen a lot of companies do this.
00:55:53
Speaker
um And then based on that status or stage in the funnel, they can do different things. They can use AI to call. They can call themselves. They can send WhatsApp messages.
00:56:06
Speaker
So like say a reminder that your interview is tomorrow ah ah if someone has to submit documents. and Exactly. Please send your PAN card. Now, so we're trying to now introduce agents, by the way, across all of these stages. AI agents, right? There might be an agent devoted just to collecting documents and validating them. There'll be an agent devoted to just reminding for interviews. There's an agent um to collect some basic information, right? So there'll be agents cutting across these stages in the the recruitment process.
00:56:38
Speaker
There might be an agent that also supports people when they join the job in their early days. So we're thinking about that that piece as well, how to make people successful post-joining. um And the the agent might be the default way to nudge people along.
00:56:54
Speaker
And there might be a fallback to the human because like I said, in this ecosystem, in a country like India, the human, I think will always play an important role. But instead of being 100%, right, can we reduce dependence on humans to let's say 30% so that we have at least 70% automation, which then makes the overall processor and ecosystem a lot more efficient.
00:57:16
Speaker
ah I'm guessing would also need a... an ops team to coordinate with these agencies? I ah i've feel like it would not all be software run, right? You would need an ops team, someone who's telling these guys that his offer was released, he didn't send a document, he didn't send a phone, whatever, stuff like that. or ah you know Not really. So know we do have a small team of account managers for our top agency partners.
00:57:46
Speaker
ah We now have over 2,000 recruiters on our platform. Our ah account management team is about 30, 40 odd people.
00:57:58
Speaker
And that's only for the the large partners, right? Like the ones who we really want to build strong relationships with. We also do help them with ah planning their businesses. We help them with working capital loans, um right? Hiring, solving problems, et cetera.
00:58:15
Speaker
So we we have account managers for our largest partners. And these account managers have like strategic conversations. They are not tactical conversations like kind of.
00:58:27
Speaker
kind of
00:58:30
Speaker
So these are largely strategic conversations, though there might be pockets of operational issues that kick in. ah We have a centralized support team oh that works on such operational issues.
00:58:43
Speaker
If there are breakages in the processes, etc our either the central team will kind of work with the the employer and and solve those. Okay. Okay. Okay. ah And I guess the beauty of this model is the most of these ah people who are connecting supply and demand, there is always that risk of you...
00:59:08
Speaker
getting disintermediated. But this will not happen in your case because say Flipkart would not want to work with a small agency. Exactly. Flipkart would only want to work with you. So that revenue leakage ah is not happening due to disintermediation. Yeah. I'm wondering, I used to run a hiring agency for a decade and The biggest challenge which I faced was attrition.
00:59:33
Speaker
um I was in the white collar space, so maybe it's different for blue collar. But ah what are your payment terms? Is there revenue leakage because of attrition? How does that happen?
00:59:46
Speaker
When say attrition, do you mean attrition of agencies or? No, people that are placing. workers. yeah Yeah. I mean, typically in white collar hiring, if the guy you placed quits within 30 days, 60 days, 90 days, depends. yeah yeah You will not get paid.
01:00:02
Speaker
Yeah. So our payouts are also linked to um the productivity of the worker in their first 30 days. um And we so we recruit a lot of people in the delivery space, right? Quick commerce, e-commerce, food, etc.
01:00:17
Speaker
And this is actually a very high churn space. Most blue collar jobs are high churn. And so we do get paid basis certain milestones that they'll hit in their first 30 days.
01:00:29
Speaker
ah And even though this is a high churn space, we still make enough money to make a margin. even after paying agencies. right So it's actually very, very... Yeah, we're all... Our entire team is about 150 people, right? And we recruit over 40,000 people a month.
01:00:46
Speaker
Wow. So it's a very, very like high leverage business in that sense. um But yeah, to your point, high volume, low margin, but there still is enough margin at scale. Wow. It's a very efficient business. And the learning was that the economics work really only at this local level.
01:01:03
Speaker
right If you try to centralize all of this, ah it becomes very, very hard to juice out margin because your costs also as a as a large organized player are much higher.
01:01:14
Speaker
But these local agencies who operate at a very, very sort of ground level are able to operate at much more efficient cost structures. Okay. So a comparable business to you would be like, say, a team lease?
01:01:30
Speaker
No, no. Teamly is a very different animal. They're a staffing business. So Teamly is now hires, I think, 15,000 people a month or something of that sort. And then they put them on their payroll.
01:01:44
Speaker
And then they get paid a commission on the salary of the worker. Very, very different business. Yeah, very different business. Yeah, very ah very traditionally run. Obviously, they're they're also doing a bunch of stuff with with technology. Everyone is.
01:02:01
Speaker
ah But a very, very different kind of business, different DNA. Honestly, there isn't any direct comparable. There are people in the space. Team Needs is one. Upna, as you mentioned, is another. There's company called Better Place, which is largely doing staffing and the background checks. There's an another one called Work India, which is also a marketplace, but more for ah small businesses to hire.
01:02:22
Speaker
um So there are players, but in this particular segment, which is, let's say, delivery, Qcomicom, we are the largest player by far. And luckily, knock on wood, we don't have direct direct competition.
01:02:35
Speaker
What percentage of your recruitment is delivery? I think today it's about 95 plus percent. Most of it is.
01:02:45
Speaker
We have recently started recruiting people for warehouses, for ah even Uber drivers and even factory workers. um So the same model, by the way, the the agency model is now kind of scaling horizontally and quite well.
01:03:03
Speaker
um the right The way I talk about it ah with my team is that we've done our books, right? Like Amazon TripCart, the first category they cracked was the books category. For us, that was delivery.
01:03:13
Speaker
And now we're trying to add other categories to our business. So ah you're saying that for delivery gig workers, you are the dominant player in this market and...
01:03:24
Speaker
There is nobody else who's really focused on delivery gig workers as such. No, people have focused. ah There are companies that are trying even today, but we are by far the largest player.
01:03:37
Speaker
So you said that for qualifying, it was 510 rupees. For giving leads, it was 1500 rupees per lead you provide. Now for placement, what is it now? like For placement in this space, it's at least like 15 to 30x of that.
01:03:56
Speaker
ah right So you make upwards of 1500 all the way to let's say 3000. And in the manufacturing space, you make... upwards of 5,000, 6,000 rupees ah per placement. So you capture a lot more value.
01:04:12
Speaker
The manufacturing pays more because those are not gig workers, I guess. Those are more like permanent hires. Therefore, the payout is better.
01:04:23
Speaker
ah Yeah, I think that could be a part of it. I think um the volumes are also... Lower attrition there also, I guess. because Yeah, the attrition is lower. I think it's perhaps a slightly higher scale job as well.
01:04:37
Speaker
right Because you're ah you're working on an assembly line, you're not just delivering ah parcels. I think it's tied to that. yes Okay. So, 40,000 recruitments a month. ah Who else does this kind of number? Like you said, Team Lease is 15,000 recruitments a month. Nobody that I know of, ah in India at least, ah ah has it touched that volume. And i mean even 40k is a drop in the ocean in this country. Our goal, hopefully in the next six months, or maybe sooner, is to cross
01:05:13
Speaker
And we do see a path towards getting there. my My big, hairy, audacious goal is to get opportunities to a billion people in the next five years. ah Not just in India, but perhaps other markets like Southeast Asia, the MENA region, etc.
01:05:28
Speaker
And I think if some of the bets we're making with the especially the AI piece turn out well, right then we can see that kind of ramp up. Okay, fascinating. out And so beyond the placement part, are you also going back to your upscaling roots?
01:05:52
Speaker
Yes, so we made ah a strategic um acquisition, in fact, of a company and a product ah called Learn, which was built by a company called Goodworker. Goodworker was, in fact, a competitor ah that came about, I think, around 2021 or so.
01:06:10
Speaker
ah They were trying to replicate our model, Temasek funded. um This came sort of a full circle where Temasek funded us through an early stage venture arm called Lemmetry. And we acquired ah this product called Learn From Them, which is a skilling product, which imparts skills through short videos.
01:06:33
Speaker
And I feel that while skilling is not a great hook, it's not a great starting point, but it's a great add-on. Because there is some percentage of this audience that would want to skill up and that would even pay for learning new skills.
01:06:46
Speaker
So this is something that we will introduce into our product stack fairly soon. You want employers to pay or the candidate to pay for upskilling? So in my mind, um there would be two models. I think and some candidates would pay for learning new skills.
01:07:06
Speaker
And I think employers obviously would pay for hiring better skill people. So I think in a way you would be able to extract value from both sides. Okay. Okay. okay ah Okay. So I want to go a little bit into the product per se, the Y spot that you're building. Okay.
01:07:27
Speaker
What is your moat in that? ah Is it a like, ah ah are you using a foundational model which is doing most of the heavy lift or is there something over and above just using a foundational model to ah power a voice port?
01:07:45
Speaker
Yeah. Yeah. Yeah. So in our case, the moat really comes from, I think, two things. One is just sort of the experience that we're trying to design for this audience. And I believe that we understand the audience very well.
01:07:58
Speaker
um And that would be important in this context because it would help bridge trust. And the second is oh the data sets that we possess. So we we possess hundreds of thousands of hours of call recordings between humans, human recruiters and human candidates.
01:08:17
Speaker
And we are leveraging that data set to fine tune some of the models that are being used for the bot. Now that is also a very, very unique a dataset. I don't think a lot of companies would have that.
01:08:29
Speaker
um And I think it it kind of goes back to ah creating, you know, verticalized AI solutions to create defensibility now in this AI-first world, right? I think AGI is perhaps not very far away and AGI will be able to do a lot of things at a horizontal level, but you need to find use cases where you can sort of, you know, go sort of inch wide but mile deep.
01:08:56
Speaker
And I think blue collar recruitment at scale in a country like India is one of those, you know, use cases where we can really, really go deep and build a lot of differentiation. This data set of calls is ah generated by the agency partners. Like whenever agency partners are speaking to candidates. we can Correct. We are able to record those calls to our product. Yeah. ah Okay. Okay. Okay.
01:09:20
Speaker
And for the agency partner, it's a free thing because they are essentially getting, like like there's a revenue pass-through. It's not like a SaaS subscription. they have to we don't We don't directly charge them today for this. kind of So they are incentivized to use it instead of doing calls on their own. And very, very interestingly, right? Like the whole AI world is now starting to move towards an outcome-based revenue model.
01:09:45
Speaker
where you charge only when ah you deliver an outcome and we've been outcome based on day one absolutely beautifully in a country like India as a founder who wants to build in the AI space like how do you build vertical AI how do you build a product moat you know because it ah what makes it hard for someone else to come and use the latest chat GPT model, create a voice word and replicate what you are already doing?
01:10:25
Speaker
Yeah. Yeah, it's a great question. And I think the answer will perhaps keep changing as the the technology evolves. But my current answer is sort of linked to what I said earlier. I think the first ah element of this is really understanding the customer or the user very, very well and understanding their pain point well and therefore building a better solution that can solve their problem better. How does that building happen? Like if you can go a little...
01:10:55
Speaker
ah technical for me. o Oh, like, do you do which ah model do you work with, for example, which foundational model? And then how do you make ah your moat on top of ah foundational models?
01:11:09
Speaker
Right. So in our context, for example, ah being able to handle different types of languages or accents or dialects would be very, very important. Right. And therefore, ah at any given moment in time, we ah you know are are trying to at least test most of the latest, for example, speech-to-text models that are available.
01:11:31
Speaker
um OpenAI has one called Whisper. Google has many of them. 11labs is also a usual suspect. And then there's something called DeepCram.
01:11:42
Speaker
And so we use our human-human call recordings to benchmark all of these systems on a continuous basis. Every now and then, there'll be a new version of these and many others, right? So we keep benchmarking them.
01:11:54
Speaker
And the the one that works the best is what we deploy in production. um And we do that for the entire stack. So we have a speech-to-text system, we have an LLM, and then we have text-to-speech.
01:12:07
Speaker
Text-to-speech is less so ah difficult, but the speech-to-text and the LLM pieces are the tough ones. right so What does the LLM do? it ah like It authors the response. It composes the reply. Correct.
01:12:23
Speaker
So it takes the text from the speech-to-text system. term It understands that and then it authors the response right that you then send back in the form of speech. So all of these systems have to come together and they have to work in fairly near real-time, low latency ah so that you can deliver that experience on the phone.
01:12:40
Speaker
right So then there's a lot of heavy engineering under the hood ah to ensure that low latency, which is really, really tough to deliver. And then cost. Ultimately, in India, for this kind of use case at scale, ah you don't want these systems to burn a hole in your pocket, which they can.
01:12:57
Speaker
So we do a lot of different types of caching across the board. I think our current voice bot costs close to two rupees per connected minute. ah which is in fact cheaper than a human now.
01:13:10
Speaker
I think human cost comes up to some, if I remember correctly, three to four rupees per connected minute. So we're were going to bring it down to lower than human cost.
01:13:22
Speaker
ah That's important as well. And then, like I said, ah bringing in the nuances of the use case through the data sets that you possess and possessing a unique data set that not many other companies in the world have access to, a very domain-specific data set is a superpower.
01:13:42
Speaker
And then using that data set to really drive strong value, ah right that's something that ah is something that's important. How do you customize a voice model? Like you said, Whisper and Gemini has some, Google has some, et cetera. How do you customize it? Like you will feed it the recordings and ah like like, again, from a slightly helping a non-engineer understand the technology of it from that perspective.
01:14:10
Speaker
So um we use recordings in multiple ways. Uh,
01:14:17
Speaker
One thing that we do is we transcribe the recordings into text and we can use that text to ah evaluate different parts of this stack. So the LLM, for example, ah based on the responses it generates, we can quickly evaluate how it's doing and make changes accordingly.
01:14:37
Speaker
ah We're also trying to use the speech data in order to fine tune the speech to text system to make that better at recognizing Indian ah languages and dialects and accents.
01:14:49
Speaker
So this data set can be used in many, many different ways. um We're in fact now trying to go so one step further and we're trying to figure out if we can predict how much the worker might retain on the job.
01:15:05
Speaker
So based on the signals that we capture on the call itself, we're able to predict, ah in fact, our early experiments have given us, ah I think, a 70% accuracy already on predicting yeah whether the person will take up the job and stay.
01:15:22
Speaker
um So that's another very interesting use of this sort of data set. ah So yeah, it's ah it's a it's a very, very wide canvas process. Are you able to ah label emotions?
01:15:38
Speaker
like Like the thing about a human calling versus a bot calling is ah the same text, ah like just even in terms of how someone says, okay, versus okay.
01:15:51
Speaker
yeah you know A human knows the difference between these two okays. Does that happen with AI yet? So it does, I think, happen at a subliminal level.
01:16:04
Speaker
um One of the things that we are building, like I mentioned, is ah is a model that can predict whether somebody will stay on the take up the job and then stay.
01:16:15
Speaker
And we've tried um a version of this where we are feeding in the raw the speech data from these calls. And that actually works really well.
01:16:27
Speaker
And these are very simple calls. They're not really complicated. you know We're just asking them a few basic questions, answering a few questions. But it seems like the models are able to capture some amount of that that nuance you know that is not explicitly mentioned.
01:16:44
Speaker
And that's why like they're working reasonably well. like like We're getting 70% accuracy with a very simple sort of model. in this way. It's a black box process. Like like the model is spitting out a number. you don't necessarily know how it's calculating it.
01:16:58
Speaker
So it could be reading the emotion. We are not sure. It could be. It could could be. Got it. Okay. that that mean That's not the hypothesis. But yeah, I mean, it's it's a complete black box. It's hard to 100% hundred percent and Okay. Okay. So, ah you know, if you were starting up today with the state of AI that is available today, what how would you go about building and finding PMF?
01:17:25
Speaker
That's a great question.
01:17:29
Speaker
What market would you select? Where do you think the opportunities are? Hmm.
01:17:39
Speaker
I think there are a lot of opportunities in the enterprise space. And I'm a little more biased towards, i think, enterprise because I've seen that the most. ah Every single large enterprise wants to do something in a eye And so every single large enterprise in the world almost has a budget to spend on AI, which is why you're seeing AI companies generate tons of revenue very quickly.
01:18:06
Speaker
ah Revenue retention, et cetera, is a completely different topic. ah But getting the paid is not as difficult. Now, retention will come from solving, like i said, a painkiller versus a vitamin um and solving it better than others, right?
01:18:23
Speaker
So I think the hard part about this is really one discovering what that painkiller is within said enterprise, let's say, or enterprise segment, and then building a much better solution for that.
01:18:37
Speaker
A better solution could be just better user experience or being able to deliver more value through some sort of maybe data set that you possess. Yeah.
01:18:49
Speaker
yeah Any spaces that you feel are currently there for the taking within enterprise? Any pains that you can see based on your interactions with your clients?
01:19:03
Speaker
I mean, so obviously recruitment is ah is a big one that we are attacking, especially for blue-collar, large-scale. I think just a recruitment in general is is across the board, white-collar, gray-collar, everything. It's such a broken process when done manually. You've probably seen this yourself.
01:19:19
Speaker
A lot of that is going to get automated. and It is. I mean, so many startups attacking that. I think um outside the recruitment space, ah anything where...
01:19:34
Speaker
There's a lot of deep human involvement and where workflows, I would say, are very media heavy. When I say media heavy, where you have to, let's say, work with a bunch of documents or you have to work with a bunch of, let's say, images or videos.
01:19:55
Speaker
So ah legal, for instance. ah Therefore the taking. Marketing. Again, therefore the taking. Accounting to some degree now also because these models are becoming better at reasoning and math. ah right So eventually you'll get to a point where ah they're able to actually solve problems. So Not only are they generating the media or maybe making some sense of the media, but they're able to do some really complex reasoning, you know, through the media. you give we did get that
01:20:32
Speaker
But yeah, i think top of my mind, those are some spaces that, you know, are are fairly hot today.
01:20:41
Speaker
So you've largely been B2B so far now with some experiments on B2C side. ah Yeah. Is there a conflict between these two where, I mean, imagine Zomato starting its own restaurants, ah you know, so so the restaurant partners who trust Zomato to give them business might revolt, et cetera. So you are thinking along somewhat similar lines. ah How do you think that will play out?
01:21:12
Speaker
Yeah, so my sense is that in a country like India, 1.5 billion people and growing and so much complexity in terms of cultural ah history and and nuance, these models will will coexist.
01:21:29
Speaker
The B2B model will be important ah in this case so that we can bridge trust with the end user. um And I think the consumer model will perhaps help us capture ah more scale faster.
01:21:44
Speaker
um In fact, to your question, I think i don't know if Zomato has done this, but Swiggy has created its own ah food brands for sure. So I think ah these models, the B2B and B2C models will coexist in a country the size and scale of India.
01:22:01
Speaker
ah Once they've gotten a job through our platform, right we can then offer them maybe their next job or other services like outskilling, etc. And that will lead to us ah just helping us scale further and you know deliver more impact faster, ideally through our AI, which today is an AI recruiter. But we see that evolving into almost like an AI advisor that can help people get access to loans so that they they can start businesses, maybe their next upscaling opportunity so that they can grow in their lives and their careers.
01:22:33
Speaker
So that's the the vision that we have for the company. Fascinating. Thank you so much for your time, Madhav. It was a real pleasure. Thanks for having me, Achy. It was great to talk to you.