0:00
I was just panicked because I never prepared for anything.
Search this video
0:01
I mean, I prepped so much for the interview that I chatted with a lot of XYZ interview founders, etc.
0:07
They all told me, "What's your idea? What's your time?" and everything.
0:11
But, Harsh didn't ask about any of those things.
0:13
And I thought genuinely interview went like so horrible that we are not going to get in.
0:16
Harsh told us, "You guys are really good engineers.
0:21
Just pick something else and work on it." >> [music] >> All right.
0:27
We're so pumped to be here with Varun from GigaML.
0:30
Um thank you all so much for coming.
0:33
Varun, why don't you start by telling us a little bit about what GigaML actually is? >> Sure.
0:39
We build AI agents for customer support.
0:41
We work with some of the biggest companies in the world like DoorDash.
0:45
We work with one of the biggest crypto exchanges in the world.
0:48
Top three telecom providers in the world. Yeah.
0:50
>> And so, what does it mean to when you say AI agents for customer service?
0:51
Give an example of how someone is using your product.
0:56
Maybe even how people here have used their your product and not even realized.
1:00
>> The traditional way that support works is whenever is called the support, it actually uh get to an IVR or a chatbot, then goes to a human.
1:10
The deflection rates are closer to like 10 to 15% with AI.
1:14
It's just like you call, it's entirely like human-like experience.
1:19
Closer to like 60 to 70% deflection rates.
1:21
We're aiming to get it closer to like 90 to 95% for top of the customers.
1:25
It's just like a better experience.
1:27
You have never need to be in hold again.
1:28
You can just call and get the issue resolved like really fast. Yeah.
1:34
>> And why don't we start by rewinding the clock a little bit and hearing about the very early days of your company?
1:40
So, maybe even before you started a company, why don't you just tell us a little bit about what was your upbringing like?
1:47
Like, how did you uh get introduced to technology in the first place? >> Yeah, of course.
1:51
Uh my uh I was from this town and small town in Andhra Pradesh and my parents were both like government teachers.
2:01
They of course wanted to be an engineer or like a doctor.
2:05
So, we were uh, I grinded myself out to get into IIT, which was a great experience.
2:11
I got into like IIT Kharagpur uh, in electrical engineering and my first 2 years I did not like do much because it was like a COVID year, which was like partying and like not much studying.
2:23
And during third year I started like uh, doing my research in LLMs in Stanford and that's when we started like uh, seeing this is like a pre-ChatGPT moment.
2:35
We're working on transformer models like BERT etc.
2:37
And >> This was like when you were in college or still in high school? Okay, cool.
2:44
>> This was when I was in college and I actually got a like pretty good job on one of the quant firms uh, one of the leading quant firms in New York.
2:53
I got offer like 50k at that point of time it was like a big thing uh, and I also got my PhD in Stanford uh, to join and then ChatGPT launched.
3:03
Uh, this this is uh, around December is when I got in and then ChatGPT launched.
3:09
We were like super excited.
3:11
Uh, the super exciting moment is it was able to write code and a lot of things.
3:15
So, we just wanted to build something on top of it and give getting into YC a shot because I was I was reading PG's essays watching like the 2014 YC startup school.
3:23
It's a classic one and I just wanted to give it a shot and try to get into YC.
3:29
So, me and my co-founder we both know each other from like freshman year and we just thought like what could go wrong?
3:37
You can just apply and always see, right?
3:38
And it's so dramatic because I was supposed to join we were supposed to like join in like 3 days.
3:42
That's when finally when we got in and the interview was also like so different than what we thought because we we were trying to build an EdTech on using like LLM and we go to interview.
3:54
I clearly remember Harj asking me that, "Hey, this is EdTech. It's not going to work. Pick something else."
4:00
And it was never even about like the idea or anything we wanted.
4:02
He was like, "You have like research experience in LLMs. Pick something and etc."
4:07
>> And so so when you heard that did that increase your ambition or reduce it or how were you thinking about that?
4:13
>> I was just panicked because I never prepared [laughter] for anything.
4:15
I mean, I prepped so much for the interview that I chatted with a lot of YC interview founders etc.
4:19
They all told me, "What's your idea? What's your time?" and everything.
4:23
But Harj didn't ask about any of those things.
4:25
And I thought genuinely interview went like so horrible that we are not going to get in.
4:31
And yeah, Harj actually wanted us to join YC and he told us, "You guys are really good engineers.
4:37
Just pick something else and work on it."
4:38
That's how I mean, we would You know, we would not have existed without Harj taking a bet. Yeah.
4:44
>> And so okay, so I imagine then you went back to your roots a little bit as engineers.
4:48
Tell me a little bit about what were your roots as engineers like?
4:52
Like when you were in college, were you like a by the book guy or were you more of a hacker?
4:55
Like what was your kind of personality as an engineer?
4:59
>> Yeah, it's it's very interesting because my co-founder was I think third ranked in entire IIT and he only got one B in his entire career.
5:08
That too he really messed up.
5:10
Actually, there is a he actually went to class and gave attendance for a wrong guy as a friend and he got penalized 10 marks.
5:17
That's why instead of EX he got B.
5:18
Otherwise, he would have been like top of the campus.
5:21
He's like a very book guy kind of a person.
5:23
For me, honestly, I never my grades were bad.
5:26
I never like really studied that good.
5:28
And I used to do a lot of Kaggle competitions primarily because you can make money if you win the competition. That's pretty much it.
5:36
So, I would say I'm more kind of like a hacker and he's more kind of like a book guy.
5:39
But eventually, we both turned out to be like a hacker kind of people. Yeah.
5:43
>> Wait, can you Yeah, can you tell everyone about your history with Kaggle competitions?
5:45
Like how many of these did you do and like how much money did you make doing this? >> I did a lot of them.
5:52
I think I made a lot like $50,000 or something.
5:54
That's how I landed to like one of the >> [laughter] >> one of the one of the high frequency trading jobs.
6:00
I gamed it so much that they banned me.
6:02
I Yeah, we we I gamed a lot of that thing. Yeah.
6:07
>> So okay, so when you were in college, you're now spending a lot of your time doing Kaggle competitions.
6:11
You're experimenting with machine learning models.
6:16
You apply to YC eventually.
6:16
Harj tells you, "Okay, this like EdTech idea is dumb. Do something else."
6:20
How do you now bridge those two worlds?
6:23
Like what's kind of the next thing you decide, "Okay, let's try to give this a shot."
6:28
>> No, it's Harj did a pretty great job.
6:31
What he did was he made us chat with Coursera COO and a bunch of people who have like some of the most successful EdTechs.
6:37
All of them told it's a bad idea to do EdTech.
6:39
So, we decided to pivot pretty much after a month into the batch.
6:44
And we were Again, we had like a pretty bad experience because both of our B1 B2s got rejected, but this was the first time YC turned into in-person.
6:53
We're just doing fully remote.
6:53
We had experience on fine-tuning, so we just I just read this research paper from one of the DataBricks co-founders, which is basically LLM caching LLMs to reduce the cost.
7:04
At the time GPT-4 used to be super expensive.
7:08
Instead of just caching, we thought that fine-tuning might be better using a small LLM and everything.
7:12
So, that's how we got started with fine-tuning.
7:14
And we open-sourced a bunch of models.
7:16
We topped Hugging Face benchmarks.
7:18
That's how we got to like a lot of traction and a lot of people reaching out to us.
7:21
And we raised a $4 million seed round.
7:25
>> So okay, so you did YC then.
7:25
You end up raising your seed round.
7:29
You're You're working on this fine-tuning idea.
7:32
That doesn't sound anything like customer service AI system.
7:33
So, how did that first thing become the second thing?
7:39
What was that What was that story like in which you found the idea you're now working on?
7:43
>> Yeah, the interesting part about fine-tuning is it's a really bad market for a bunch of reasons.
7:48
The only reason you want to fine-tune is to reduce the cost and make it faster.
7:51
Or else there's another use case which is if you were so secure, it's very hard to sell to those big insurance companies or health care as an engineer because it's a sales process, not engineer process.
8:03
We have realized it after like a year or so.
8:07
And the interesting thing was the only two use cases on GitHub customers which are growing very well are customer support and coding.
8:14
So we decided to customer support and >> So so you basically saw that from your customers, right?
8:18
It wasn't like you just like looked and said, "Oh, this is a big market."
8:23
You just you naturally discovered it from the thing you're building.
8:27
>> Yeah, we saw it from customers and I was just here and Zepto was a first customer for that thing.
8:31
We just reached out to them and as they were scaling really fast and they tried us out and they turned out to be our first customer.
8:39
>> So okay, I mean, this is where I think things can be very counterintuitive, right?
8:42
Like when you chose to switch into customer service, there existed at least one or two other companies, some of which who have very famous founders, who are very well capitalized, companies like Sierra for example, that already existed at the point of you guys deciding to go all in on this.
9:00
Why did you think that you could do it anyways and still win?
9:02
Why did you think that it didn't matter or it was worth it for you to go after that market?
9:08
>> Firstly, we don't know CRN account existed when we signed Zepto and >> So like there was a little bit of benefit to naivete there, yeah.
9:16
>> That thing and we didn't think of competition much.
9:18
That's pretty much what I would say.
9:20
We never I mean even in right now our entire product is is is a is a are the customer willing to pay you?
9:25
Can you deliver a lot of value to them?
9:27
Just like do it was like a pretty much a stronger mentality against competition.
9:33
But again, as you scale, you need to do differentiate and everything.
9:35
Then after Zepto, the real competition happened between us and one of the leading companies that you were saying at DoorDash.
9:42
We were like eight people going against this 400% well-funded company, and we won it against them.
9:46
And as you know, DoorDash is one of the massive support levers.
9:50
That's when we truly like us realized that there is a lot of arbitrage of actually building a great product rather than a sales team.
9:59
>> So, okay, so that's kind of crazy, right?
10:02
You you won DoorDash's contract.
10:03
You were only a team of eight people when you did that.
10:05
Why was it that DoorDash, which is a huge company, was willing to trust you guys when you were only eight people?
10:11
You could imagine they're like, "Oh, these guys are too small.
10:15
They'll never, you know, they'll never trust me."
10:16
Like, you know, 10 years ago, the assumption would have been that big enterprises would never buy software at that scale from a startup.
10:23
It seems like that was no longer the case for you.
10:24
Why do you think that is?
10:26
>> I think uh you got to have some unfair advantage, right? For us, YC was that.
10:30
Garry intro'd me to uh Tony and he's a YC company, and DoorDash is also basically a YC company.
10:37
So, it's just like an inherent trust to it.
10:39
And again, it's I mean, we piloted for like 3 months.
10:41
We never went down, and all the metrics were good.
10:42
And DoorDash is a very meritocratic company, and I really kudos to them.
10:46
And it's a the company at that scale picking a small company would be hard.
10:50
And now, a lot of companies pick us because of DoorDash and a lot of other big public companies use us. Yeah.
10:57
>> So, how is that What does the article look like since then?
10:59
So, you got companies like Zepto and DoorDash in the early days.
11:02
What does your company look like now, and how have you had to evolve what you're doing over the last uh few years?
11:09
>> Yeah, last few years uh I would say we as I mentioned to you, we work with the biggest crypto exchange in the US.
11:16
And after getting like a customer at that scale, we uh we are working with a lot of Fortune 500s, etc.
11:21
, like uh trying to automate the support.
11:23
Another big interesting thing that I have observed with AI agencies is fundamentally boils down to two things.
11:30
This is true for almost any identity company.
11:31
It's like a policies are the markdown file and how can you iterate the markdown file to affect a business KPI?
11:37
That is support for support resolution rate or like CSAT.
11:43
The all thing that matters is you let's say we started like 30 to 40% resolution rate, how to get to 90%?
11:48
How can you iteratively improve to get that?
11:50
The same fundamentals apply for compliance, ITSM, ITSD and everything.
11:53
So we're seeing customers, some of the biggest consumer companies in the US piloting us for internal support, even compliance and etc.
12:04
It's very interesting with AI agents.
12:06
It's fundamentally marked down to markdown and like how can you iteratively improve markdown to move a KPI?
12:12
>> Okay, so I want to change the change gears here a little bit and talk a little bit about the advice you would give for this people in this room, right?
12:21
So the people in this room are predominantly college students or young people recently out of college.
12:24
You know, if you were rewinding back to your time as a college student, is there any advice you would give that's maybe different from the advice that they might be hearing from their peers or what is kind of the norm, you think?
12:39
>> Yeah, I mean like a lot of people thought I was stupid when I rejected a job offer.
12:43
I mean like it was so insane because uh >> They were like you're turning down this great offer from a quant firm.
12:47
Like why would you do that? Yeah.
12:51
>> It's just for me and my co-founder we just wanted to give it a shot.
12:53
We thought like we just wanted to see how how how high we can go.
12:57
That's pretty much like a thing and that's the reason that's the pretty much the entire way that we built the company.
13:05
Like yeah, scaling like even denying acquisition offers from some of the biggest companies in the world.
13:09
That's how we pretty much like built the entire company.
13:12
We just wanted to reach to our potential and see is this the best we can do and push it.
13:16
It's kind of more than me, it's my co-founder's mentality.
13:20
He's fundamentally like a zero motivated by money person.
13:23
So yeah, that's That's it came from and I read Paul Graham's essay on how to do wealth.
13:30
It essentially boils down to doing a company and having or like doing equity in something big. It's pretty much it. Yeah.
13:37
>> What did like your parents think about that in those early days? Like were they excited?
13:41
You You said you know, you grew up in a very modest upbringing.
13:45
Presumably getting this high-paying job would have been life-changing for your family.
13:48
Like how did How did that go? That conversation.
13:51
You're like, "I'm moving to SF.
13:54
Um I'm going to start this crazy thing."
13:57
>> My dad was super mad, to be honest.
13:57
And it's just a I mean, it's a it's a it's kind of like a big fight thing in home, but it's fine.
14:06
I mean, like they're like, "What can they really do?"
14:07
They can't like force a kid to do something.
14:08
But it was it was just a really like a hard conversation.
14:12
I actually showed them what YC is, showed them YC videos, and showed them these are the companies that they do.
14:18
And I showed them like even if I don't succeed in a year or two, I can just go back to the job and do the thing again.
14:24
They come from like a middle-class family or upper middle-class family.
14:27
It's just like a lot of expectation to do things.
14:28
And especially when I got like a really good job offer, and my parents were all happy about.
14:34
They I mean, it's it's just not my parents.
14:38
Even internally, even I felt am I doing something wrong?
14:39
But again, my entire philosophy has been like just take a shot and see if it works out or not.
14:46
>> So, when you think about now how college students or young people can put themselves in a position to find great startup ideas, what are some of the lessons from your experience that you think apply to the people in this audience?
15:02
>> The biggest thing we made at least a mistake even after getting into YC is uh we worked on a lot of stupid ideas, which didn't make any revenue or doing anything uh for a long time.
15:14
I mean, like people just want have like a lot of ideas.
15:19
I can just go into ChatGPT and get like 10 ideas on what to do, right?
15:22
Uh it's never about the idea.
15:22
It's about if somebody is willing to pay you money for it.
15:26
And is I mean I don't even think about it.
15:29
You need to care about market, to be honest.
15:30
Is somebody willing to pay real money if you solve the problem for the value that you delivered?
15:35
That's the strongest approach we took.
15:36
And I made a lot of It took me a lot longer to realize.
15:41
Right now, even for our new products that we build, we make sure that the customer can actually pay it.
15:46
And we predict like this is the amount you're going to pay and get a commitment from the customer and then go and build it.
15:51
>> And why is it that, you know, charging for a product so early is useful?
15:53
You know, I think a lot of people in the audience might think, "Oh, I'm just a college student.
15:58
No one's ever going to pay me for something I make.
16:00
Right now, I should I should just get started for free or something like that."
16:03
Why does that not tend to work in your mind?
16:08
>> If it's an important enough problem, people would pay.
16:09
Either with money or with time, I would say.
16:11
I mean, social media networks are just time.
16:13
They don't charge you any money.
16:14
But in general, for any single B2B company, if the problem is important enough, people should be willing to pay money for it.
16:22
Otherwise, like you're just solving a fake problem.
16:25
>> So, you are also part of a new generation of Indian origin founders who is building your company both here and in SF.
16:34
How do you think about that?
16:34
I mean, do you think about that or is there a any any particular framework you have for how to expand your company and take advantage of your background from India?
16:45
>> I think that you should just stay close to customers wherever you are.
16:49
But if you're doing anything closer to like GenAI and very like research-based things, I strongly think SF is a place.
16:57
Because the amount of access as in uh with the researchers and etc.
16:58
you get is insane because almost all of the innovation in this GenAI field is getting drived on Bay Area alone compared to India.
17:07
But again, if your customer is primarily based out of India, you should be here.
17:12
>> What do you think when it when you look forward to the future and you think about, you know, all of the pivots your company has gone through to get to where it's at now, how do you think what do you think the next few years look like for your company?
17:24
Like, do you see yourself now do you feel like you found the thing that's going to become really big or do you see yourself continuing to evolve and how are you evolving yourself?
17:33
>> Yeah, I think I think we're I'm very confident that we're moving in a great direction.
17:37
The biggest bottleneck of every single enterprise AI deployment, regardless of support or any automation that I want, this concept called forward deployed engineer, you need to have them bunch of them coming and sitting with the customers and configuring it.
17:51
We're trying to build an AI forward deployed engineer.
17:54
We're going to launch it soon. >> Yeah.
17:56
>> And that is where I would say that I'm very confident on because whenever you want to make these policy changes, whenever you want to like just spin up a new dashboard for me or how can I go from like 40 to 60%?
18:07
Our AI forward deployed engineer is going to join is on Slack, is going to join Google Meets and take all the notes and do the changes automatically. >> Yeah.
18:15
>> So, I'm very confident on the direction that we are moving on how we are going to tackle AI adoption in enterprises because the biggest bottleneck right now is forward deployed engineer and we're going to take it over.
18:28
>> Could you tell us a little bit about how your company actually runs internally using AI also?
18:32
Like you obviously your product is very AI driven, but in terms of the company itself, like what are the tools that your engineers and sales people and what not are using?
18:40
What does that actually look like?
18:43
>> It's funny because we have a one of our values is automate, automate, automate.
18:47
So, we force people to like use as much automation as possible and the in general the overall mission of the company is to automate all of the world's work.
18:55
We are intentionally like moving in that direction of generic automation builder able to automate anything on top of us.
19:01
Some of the examples were you don't need to have a personal assistant, just add it.
19:04
Open Clouderons and schedules the meets for you to actually like sales people use it very interesting.
19:11
They pull transcripts from Gong to do analysis of what are the biggest things that work are working against a specific competitor across like a multiple things.
19:21
So, yeah, everybody's That's the one thing I love about Claude code.
19:26
It turned a lot of people into builders.
19:27
And people are using it innovatively to drive specific insights that would have like taken like humans go through a lot of things and doing it.
19:36
>> Yeah, I mean, if coding agents didn't exist, how many engineers do you think would be at your company versus what it currently is?
19:42
Like how much What What would the company look like without all of that?
19:45
>> I think there should be at least like six to seven times more compared to now.
19:49
We're relatively very small engineering team.
19:52
It's like a very strong talent that's on very small engineering team. >> That's crazy, right?
19:55
Like 7x the number of engineers you would need if you didn't [snorts] have one tool but probably cost you much less than seven times your engineering team.
20:03
>> It's It's It's more than cost.
20:03
It's just It's better without context switching.
20:08
It's better for you to own the thing and build the entire thing rather than having like a lot of people working on it.
20:13
You can just ship much faster.
20:13
And the context transfer actually kills a lot of things and slows down things.
20:19
>> So, given that, I imagine the kinds of people you look to hire on your team also are very different from the average SaaS company 10 years ago or at least in some ways different.
20:28
I'm curious how you would describe that.
20:30
Are there things that you guys look for that are maybe different than what like a big tech company who doesn't use that many that much coding agents yet might look for?
20:40
>> Our interview process specifically designed We ask people to write code and remove AI and ask them to like change the code.
20:47
>> You actually have them write coding in your interview process?
20:50
>> We ask them to write code and then remove access to the tool and ask them to change the code without AI.
20:54
It's intentional because we want people to understand the code as well on how it works, etc.
20:59
So, we are I mean, again, this is also keeps evolving as AI models keep getting on better and better.
21:03
Do you really need not to know how the code works even if Claude does the entire thing?
21:09
But, that's how That's where we landed on.
21:11
Generally, the things which we look are some sort of like extraordinary ability and spikiness that we can like relate into.
21:17
For me and my co-founder, that for us was me having the highest one of the highest offer jobs.
21:23
My co-founder was third in the entire IIT and um he also got the I think the highest paying job offer in India in a in a quant firm.
21:30
And we're trying to look like a very spiky things on which is that 0. 1% of people would do.
21:37
And that's how like we're generally going at things.
21:40
>> So, you know, I think a lot of people might look at themselves and say like, "I'm a technical person. I'm an engineer.
21:45
But like, what do I know about business? Right?
21:47
Like, I don't have a business background necessarily."
21:50
How did you guys think about that, right?
21:53
You and your co-founder are both computer scientists.
21:54
I assume you did no business before you started this company.
21:57
What has been your experience on whether that that pre-exposure matters, whether if you just kind of learn it as you go?
22:04
>> I mean, there are a lot of people who will buy your product without business background.
22:07
You just got to find the right buyer.
22:08
I mean, like Zip is one of those companies.
22:10
They didn't care you have like a bigger sales people.
22:13
DoorDash doesn't care about sales people.
22:14
You just got to find your ICP.
22:17
>> But I guess I mean more for you guys as founders.
22:19
Like, in your own founder ability, do you feel like it would have benefited you or do you feel like your technical abilities allowed you to get somewhere in itself and then, you know, that's what kind of gave you the opportunities?
22:31
>> I I personally like bias a lot towards builders and sellers because uh this is the mistake I made while starting the company.
22:38
Me and my co-founder had this huge debate because I thought sales was the most important thing in the company.
22:43
I was so wrong in like so many ways.
22:43
If you take a look at all the successful AI companies, it's product.
22:48
None of them are succeeding.
22:50
I mean, nobody uses Anthropic for the best sales team.
22:51
I don't think that I mean, Anthropic and OpenAI doesn't even pay sales people commissions.
22:56
That's how they don't care about like a sales.
22:58
And I think with AI, product is the most important thing.
23:02
How how how good is your product at delivering a lot of value to the customer short amount of time.
23:07
If you can prove that, everything else should follow through. >> Yeah.
23:11
Okay, we're almost at time here.
23:14
I'm curious if you have any parting final thoughts or reflections.
23:16
You know, you've now been at this for two or three years.
23:20
You're no longer in school.
23:22
You've been spending your time split between San Francisco and Bangalore.
23:24
Are there things that you now realize about how the world works?
23:29
Um that maybe weren't obvious to you when you were getting started and you can leave people with some parting advice.
23:36
>> Yeah, the biggest thing is just getting started on trying to sell the thing.
23:38
I just getting started and jumping the thing and burning the boats is a good thing to start.
23:43
It it is only like really valuable and things get really real if you burn the boats.
23:49
That's when I really felt because we know that when the company was not working, me and my co-founder were thinking, "Oh my god, we rejected all these job offers.
23:57
What are we going to do?"
23:58
It actually forces you to make things.
23:59
And I think it's not like really really burning boats per se.
24:01
If you have a job, you like can get a job.
24:03
It's not going to go anywhere.
24:05
But in general, just like going at it and like doing things has like a lot stronger value.
24:10
Especially with AI, it's just the cost of building things is so low.
24:15
People should just build things and try to like deliver as much value as they can to very small sort of customers and see if they can pay them money. >> Awesome.
24:23
Thanks so much for coming, Varun. Really appreciate it. >> Thank you. >> [applause]