Giga: The AI Platform for Enterprise Support

Harj Taggar 0:00 [on-hold music] Today I'm here with Varun and Esha, co-founders of Giga. Giga build AI agents for customer support and are used by some of the largest companies in the world, like DoorDash, to handle millions of customer support calls. Varun and Esha, thanks so much for joining us.

Esha 0:19 Thanks, Hutch.

Varun 0:20 Thanks. Thanks for having us here.

Harj Taggar 0:21 This is a very competitive space, AI agents for customer support. Um, there's obviously very well-known and well-funded companies in the space, like Sierra. So how has Giga, um, been able to stand out, uh, and get an amazing customer like DoorDash?

Varun 0:35 Firstly, st- starting off with, like DoorDash is a extremely meritocratic company, and that's one of the massive

Varun 0:40 volumes where we had to compete against like 20-plus vendors, uh, to win a logo like that. The fundamental approach which differentiated us from vendors like Sierra is we built product and they took a m- basically more of like a consultative Palantir-type approach, which takes a lot of time to go live. If you consider DoorDash scale, hundreds of... millions of calls, where if you say to build use case by use case, it takes months to get it to solve the entire thing. For us, we primarily built a product which makes it very, very easier for them to go live. For example, compared to other vendors where we piloted, it took them a couple of months to go live. We went live in a week. That's how fast it is, and that's one of the major reasons. And Esha can add more along the complexity side of parts.

Esha 1:22 Yeah. Just, just, you know, like, just so you know, the-- although like customer support seems like in a fairly easy use case to solve, uh, but for companies at a scale of DoorDash who have millions of calls, there are decent percentage or even like tens of percentage of calls where they are extremely complex, where you have to coordinate with like multiple parties over a real-time call. And that is something none of the other products seem to have. Like, they didn't seem to support it, and we were the only ones who built the product in a way which is very general and can also solve extremely complex use cases at scale. We are live with some of these use cases right now. Um, so majorly, it's just like a purely product play.

Harj Taggar 1:59 Many other com-- AI companies take this sort of forward deployed engineer consulting approach, um, where they're almost building some amount of like custom-

Esha 2:07 Yes

Harj Taggar 2:07 ... software per customer to get it working. But you've been able to build something that mostly worked out of the box, even for DoorDash's scale.

Esha 2:15 Yes. I think that we have a rule within our company for the product team. It's like you don't build anything custom for any customer. Everything has to be a part of the core product. So you build something

Esha 2:25 for DoorDash, every single other customer should be able to use the same product feature. So when we get like very complex use cases for some of these customers, we don't build it out separately. We just figure out, how can we scale our product to support this much more complex use case? Now every one of our customers can also use the same feature for their more complex products as well.

Harj Taggar 2:45 The reason I feel like most companies take those Palantir forward deployed engineer model approach is because it's essentially too hard to do that.

Esha 2:52 Yes.

Varun 2:53 That's the right side.

Harj Taggar 2:53 There's so much like custom work you need to do per customer-

Varun 2:56 Yep

Harj Taggar 2:56 ... to get the AI working consistently like 100% of the time. Um, how have you been able to solve that challenge? How are you actually able to have a product that generalizes across the enterprise?

Esha 3:06 So this is a bet we took a few... I, I would say like pretty much like eight to nine months ago. Um, this is something like I think very other few companies actually do it. So Python and programming is the fundamental way to build anything you want. And one of the approaches we took is that Python is a first-party member within our product, in the sense that you can write Python within the product to do stuff. And the bet we took was AI is going to write so much code down the line. AI is gonna get so good at writing code, you don't really need to write that much code yourself. And that's paying off like im- immensely for us right now because someone from operations team on DoorDash can come in, talk to our AI, and write code effectively themselves, and they just need to put in the business outcomes, and then AI is gonna write code for them and build like a very, very complex use case without like much involvement from engineers. And since you can write Python within the product, obviously there are like very good primitives on where can you write, where can you not write, where can you inject custom Python code? Since the primitive is like literally Python, it makes our product by default extremely general, and it makes it very easy for us to add new use cases because we just need to develop protocols on top of these.

Harj Taggar 4:17 Okay. So that's interesting. So it's almost like the forward deploy engineer model is basically there's business logic that lives in the head of the customer's ops team or whoever. They have to-

Esha 4:28 Translate

Harj Taggar 4:29 ... translate it. It's basically the forward deployed engineer's role is to like translate that business logic into code.

Varun 4:34 Code.

Harj Taggar 4:34 But you've actually just got AI doing it.

Esha 4:36 Yes.

Harj Taggar 4:36 And so it's like the business logic is natural English. They-

Esha 4:39 Yes

Harj Taggar 4:39 ... they're typing it into your system-

Esha 4:41 Yes

Harj Taggar 4:41 ... and your system is turning it into Python code that makes sure that the underlying product works perfectly for each enterprise customer.

Esha 4:47 Exactly. And it need not be Python as well. We convert, like for DoorDash, we convert tens of thousands of lines of JSON files, their internal JSON files, into actual natural language, w- interpretable instructions and also Python code based on like their stuff. So it's majorly like we optimized it to an extent where even the FD part of a sales process is done by an AI. So it's part of our product

Esha 5:10 instead of like a human actually doing it end to-

Harj Taggar 5:12 Yes, you do, you do have a forward deployed engineer. It's just sort of like an AI forward deployed engineer versus a human one.

Esha 5:16 Yes.

Harj Taggar 5:16 That's fascinating. And you were working on even-- As I understand it, even when you were at IIT as undergrads, you were working on fine-tuning of LLMs.

Esha 5:24 Varun did like research internship at Stanford for training these like LLMs back when like even ChatGPT-

Harj Taggar 5:31 Yeah, this was pre-ChatGPT launching, right?

Varun 5:33 Yeah. It's a-

Esha 5:34 Like-

Varun 5:34 It's like-

Esha 5:35 Transformer models

Varun 5:35 ... it's not... Transformer. It's not, not, not really a lot, but yeah. Roughly, yeah.

Esha 5:39 Yeah. I think like, you know, our experience has majorly been around like fine-tuning these massive models. Like, we were the first people to extend Llama 2's context length from four to 32K. We outperformed Claude 2 and everything. That immensely helps us because we know exactly what AI can do and where it's gonna not-- where it's gonna like, you know, just blow up. And we will only automate the parts where we know AI can do really well. And especially when-Then there are some of our products which are extremely token-intensive, and that's where you need to optimize for per token cost, and that's where, like you know, all of our previous experience also helps immensely for getting maximum per, per let's say dollar.

Harj Taggar 6:17 So going back to the DoorDash contract, so the, the speed in sales matters a lot.

Esha 6:22 Yeah.

Harj Taggar 6:22 So it sounds like a big part of the reason you're able to win a huge contract like that is your time to show them the product worked was a week versus months.

Varun 6:29 That's definitely one of the ma- massive reasons. Other massive reason is just complexity of the use cases.

Harj Taggar 6:34 Yeah.

Varun 6:34 That one thing which I want to say is, I mean, this is a very common use case in DoorDash which happens a lot, which no other vendor in the market was able to solve, which is let's say if we're here at YC office, we ordered food and you went to your home, and it, it happens a lot, people order food with, like office address at home and home address in office because they just keep switching. And you ask the Dasher saying that, "Hey, can you deliver it to my home? Because at my home it's on the way." A lot of times, Dasher being the nice people, they deliver it to your home, and then they won't be able to mark it as delivered because it will be triggered by the fraud system because they're outside the geo fence.

Harj Taggar 7:07 Hmm.

Varun 7:07 Because they're not at the office, right? So then they make an outbound call to our AI agent asking, "Hey, uh, can you mark the order as delivered?" This is our AI talking to them. Firstly, our AI agent does this. We, we check a chat history between you, the customer, and Dasher to see if you actually requested the address change. If not, while the Dasher is in the call with our AI agent, we make an outbound call to the customer in parallel. They're like, right now there are two calls running. And we ask you, "Hey, Harj, the Dasher is saying that he requested a new address. Did you actually request it, yes or no?" Once you say yes, we go back and mark the order as delivered for the Dasher. This, this is, uh, one multi-party communication and we... It's live right now in DoorDash. So these are some of the complex use cases which we solve. That's one of the re- That's one of the primary reasons that we are able to win over a company like DoorDash, is primarily ability to solve very complex use cases, which stems from ability to write Python.

Esha 7:59 Yeah. It's like, you know, solve this extreme... I think, like, when we were starting the pilot, they gave us this use case. I remember looking at it, I was like, "Oh wow, this is so interesting." Um, and then initially, like, our product was able to do it to a decent extent, but there was not, like, really very

Esha 8:15 strong, I wouldn't say, like... It would not be, like, the best experience. But we changed our product a little bit and there's, like, a new protocol where you can make outbound calls from within a single agent, so basically like a sub-agent, if you think. And then, like, you know, the experience becomes much more, much more better, and then, like, we are the only one who can do it, like, at such scale with this

Harj Taggar 8:34 And in this case, the com- the complexity in that example is that you've essentially got, like a sub-agent or two calls happening in parallel and they both need context on, on the-

Esha 8:42 Yes

Harj Taggar 8:42 ... what's happening on each, in each call.

Esha 8:44 Pretty much. It's like, you know, "Harj, like, did you actually receive your burrito you ordered to YC's office?"

Harj Taggar 8:49 Yeah.

Esha 8:49 "Because the Dasher is claiming that he delivered it to your home. Is that actually true?"

Harj Taggar 8:53 And how does it work? Like, so once... If you call me and you verify that I actually wanted the burrito delivered there, then once that call's completed, it tells the other agent that's on the other call-

Esha 9:03 Yes

Harj Taggar 9:03 ... that what happened-

Esha 9:04 Yes

Harj Taggar 9:04 ... and then that agent knows to tell the Dasher-

Esha 9:06 Yes

Harj Taggar 9:06 ... "Hey, this is totally fine."

Varun 9:07 And actually marks the order as delivered. Yeah.

Esha 9:09 And then the, the funny part here is, like, previously, yes, humans used to do it. It was a wo- worse experience with humans because humans cannot stay on call with two people at the same time. Now, like, while we are on a call with the customer, the Dasher can still talk to the AI. They can still ask about their pay. Previously, humans were like, "Okay, I'm gonna put you on hold. Let me talk to the customer," and then put the customer on hold, talk to the Dasher.

Harj Taggar 9:31 Yeah.

Esha 9:31 So it's basically a better experience.

Harj Taggar 9:33 This is, this is... I'm smiling because this is sort of like the, um... It's sort of like the enterprise customer support version of Her, [laughs] the movie.

Esha 9:39 Maybe, yeah. [laughs]

Harj Taggar 9:40 [laughs] Like the same agent talking to all these people at the same time. Are there any other ways in which the AI agents are actually better at customer support than human agents would be?

Varun 9:50 Yeah. The first start pointing is, like, we all hate hold times. I mean, like I remember we have been hold for six hours for IRS, which is like nuts if you do six hours. I mean, yeah, firstly, hold times go away. That's like the biggest thing for all the thing, all the AI agents. And the second thing is that resolution times are, like, much faster. You get to the resolution faster, it's again saving you a lot of time. Multilingual is a very major thing for a company like, uh, uh, DoorDash, and understanding multiple accents. Multilingual with multiple accents is very, very tricky because a lot of even people with not native language try to speak in very broken English. Now you've got to understand the entire thing. I mean, yeah, that's, uh, those are some of the major benefits that we saw, which is, uh, fundamental. Like the CSAT also went, like, significantly up for the, for the Dashers. Yeah.

Harj Taggar 10:36 Okay. So now you've raised this huge Series A. Um, w- why have you raised a round right now, and what do you plan to do with the money?

Varun 10:43 Yeah. It's primarily to just deal with our demand. We have, like, a lot of customers on the pipeline and a lot of massive Fortune 500 companies that are piloting with us, and we need, just need to serve them and deliver, like, high quality experience for all of their customers as well. That means we need to hire faster and scale faster.

Harj Taggar 10:58 You actually have one of probably the strongest sales pipeline of any startup I've seen. [laughs] Um, how has that happened?

Varun 11:05 Firstly, our contracts, as you know, are, like, very massive compared to, like, a lot of startups on our scale, and that means that we generally tap to, like, C-level decision maker. Once they like the product really, really well, it generally, like, cross travels a lot and they refer it to the other C-level execs who are friends. So I mean, if it's, it's Fortune 2000, how many people? There, there should be like 2,000 people if you hypothetically think of. If every one person refers you to 10 people, it should be, like, pretty easier for you to, like, get a warm network to all the intros. I think that's the primary reason. I mean, ours is a very... I mean, if you think of AI in general, the only impact it can make is, like, customer support or coding are the real use cases, uh, which we are seeing in the market. And we are becoming like sort of like board's answer to, like, what's AI strategy in this company, and

Varun 11:45 Giga is becoming one of answers. I think that's the reason which we have, like, a stronger pipeline to another anything else.

Esha 11:50 And I think like, you know, obviously getting into the door is one thing, but once you go there and talk to these people who are, like, C-level execs, they are obviously very curious about the product. And the results we have, like, you know, delivered for DoorDash and also, like, you know, the...The innovation we have on the product, even though they're not very technical people, just talking to them about it gets them so excited. "Okay, this is, this is going to transform my business." And that's the excitement we see, and that's how we just like push-- get pushed on.

Harj Taggar 12:14 Yeah. I think you guys are just the, the epitome of what's really unique about AI startups right now, is that by having the technical expertise that you do, you're able to build the product that actually works, and just the product that works and is enough to get a giant company like DoorDash as a customer, and then that's gonna spread via word of mouth and open up all these other Fortune 500 companies that want to work with you. And I just don't think-- I think that was very hard to do in the past, because most people who get the product to work reasonably well or at least well enough to close the sale, um, uh, but now we're in this new era where if you don't have the skills, technical skills that you two have as founders, you actually just can't get the products to work.

Varun 12:53 It's definitely a mix of market plus great product, I would say. As you know, all the companies are thinking what AI that they need to do, and customer support is one of the massive things of them.

Harj Taggar 13:00 Yeah. Why do you think customer support has taken off and become such a competitive space?

Esha 13:05 As you know, like, you know, currently AI models, like, you know, the recent ones, they can get gold medals in IMO. They can get, like gold medals in IOI. And the biggest challenging part there is like, you know, just the context. Let's say any model can get like gold medals in IMO, why cannot it do like the job of a manager in an enterprise? The biggest reason is it doesn't know everything the manager knows. But if you think about like programming and customer support as verticals, a customer opens up a chat with you or some conversation with you, you'll learn, learn everything about the customer in context and you will solve it. So the context scope is fairly narrow for customer support and coding. Coding is also kind of similar. You learn, you look at all the files, you reason, and then you come up with an approach. Obviously, like AI is here to just make people more productive and optimize a lot of these enterprise processes. Customer support and programming, according to us at least, is are the ones with like-- which are really fit well to deal with the current context limitations of the current LLMs. Um, and it's just gonna get, like, you know, as context gets better and everything, these models are gonna do everything.

Varun 14:08 Yeah. And people try-- I mean, right now, people are like even pr-trying to proactively give more context to LLMs as well, even that also like solves the-

Esha 14:15 Yeah

Varun 14:15 ... solves the issue and more and more use cases get unlocked. But customer support is definitely one of the-

Esha 14:19 Yeah, customer support is like in that sweet spot where, yes, you need to know a lot of the policies and everything. Yes, you need to know a lot of the customers, but it's not like too much the people, people problems and things like that. It's not just too much for the LLM to just get confused about.

Harj Taggar 14:33 Yeah. The reason we want to spend so much time on just like the product and the AI for deployed engineer stuff, this is gonna be gold. Like the-- Especially the audience that watches this stuff is very, very deeply interested in this kind of stuff, and we've spoken so much on our content about like the full deployed engineer model, and there's so much interest in that. So AI full deployed engineer is the first, like I don't think anyone's ever heard that before.

Esha 14:54 Yeah. It's like if you're building an AI product, you need to build your own FDE.

Harj Taggar 14:58 Yeah.

Esha 14:58 Yeah.

Harj Taggar 14:59 You guys, basically this

Harj Taggar 15:00 is gonna position you as like at the cutting edge. Like it's basically like everyone else is sort of doing the human full deployed engineer part, and you're already like leaps and bounds ahead because it's the AI is doing the full deployed engineer part. So Giga is 20 people today, um, and you're gonna be hiring aggressively. How do you think about the culture you wanna build as the team grows?

Esha 15:16 I want to start off with saying that not everyone is fit to join us. Um, and the people we hire, they need to be like extremely smart and very mission aligned. And as an engineer, like how I would like put the role as f- in, in our company is that we are one of the only companies, or maybe we are one of the few companies where you can have direct impact on potentially like billions of people if you join us. We are here to transform like, you know, B2C enterprises to a massive degree, and it can-- it, it's starting through customer support. We're gonna expand into other verticals. We will transform a lot of these experiences a lot of consumers have with these B2C companies. And the direct impact, let's say we go live with one of the big, like Fortune 500s, which have like billions of customers, and we automate support for it, you are basically making the experience better for billions of people. And that's one of the ways, like, you know, they need to be very impact driven. They need to have the drive to make a lot of these experiences really, really good.

Varun 16:13 Yeah. It's, it's primarily a little bit of like a very high agency and very raw IQ is w-what I would put at it. And right-- we have like some of the amazing people who are joining us. We have literally convinced a guy to not join OpenAI and Anthropic, who were paying a, like ridiculously high amount than us to join Giga.

Harj Taggar 16:29 Yeah. That says-- As a startup, even one that's growing as quickly as you are, um, how do you convince people to join you instead of joining Meta or OpenAI and the giant amounts of money that they pay?

Varun 16:39 It's primarily directly because the amount of impact. See, the-- whether this was a backend engineer and one frontend engineer who were like getting competed by-- had both competing offers. The thing was, the impact that he's gonna make on any of these companies, I mean, OpenAI is a massive company, it's not a small company anymore.

Harj Taggar 16:55 Yeah.

Varun 16:56 And the impact he can have at Giga can d- the code literally which he did writes can be transformed to hundreds of millions of people. Potentially every customer of DoorDash will get a s- sim or like a, any one of these massive customers who work with us will get di-get the direct impact. A lot of people who are not motivated by money, who want their work to be used by a lot of people, get motivated, uh, to join Giga. Again, but this is also not gonna last too long if it become thousand people. But again, right now is a sweet spot where you build something which impacts a lot of people. [laughs] Our primary mission of the company is, uh, to build a world towards perfect execution, and we'll try to optimize every single thing in these massive enterprises and try, try to optimize every process and e-every operation process and make it more and more efficient and execution driven. Uh, that's one of the biggest missions, and we will-- So we're starting with support, and we'll go to every single OpEx heavy thing in these massive companies, and we'll try to optimize it.

Harj Taggar 17:46 So you think you'll go beyond, um, customer support at some point?

Varun 17:49 Yes. We're already like, customers are literally like asking us to like, uh, it goes something like this, right? Some big of-- What are the biggest OpEx things in the company? Uh, support, where they have like a lot of, uh, hundreds of thousands of people, and then there's compliance. There's like a bunch of other things in the, in these companies, and they're literally asking us, "Can you guys-- Since you're doing this, can you also do this?"

Harj Taggar 18:08 Yeah.

Varun 18:08 So we're already seeing a lot of interest, uh, in, in these things. And one, one more interesting category of people which we have seen work out for Giga very, very well are former founders who just want to have high agency and a lot of impact.

Varun 18:20 We already have these customers. We are already running pilots with them. We can-- They can work and directly work with the end customer and build a massive impact. That's one more category where we are seeing a lot of success who are-- of the people who are being really successful in Giga, former founders.

Harj Taggar 18:33 How about we talk about the two of you a little bit, actually? I'm curious about, um, tell us a little bit about your founding story. Like, how did you two meet each other, um, decide to start a company together, and give us the origin story.

Esha 18:44 So we both know each other since like 2019 when we both went to IIT Kharagpur. Um, and then, like Varun is like really passionate about startups, I think like since he was like six or seven, and he has been reading Paul Graham essays, Elon Musk essays, and all of that. Um, and then he has been trying to do a lot of startups during the college, and then we eventually a-applied to YC in 2023, and you were our interviewer.

Harj Taggar 19:08 Yeah, you applied with a completely different idea. [laughs]

Esha 19:10 Applied with an-

Varun 19:11 Oh my God. [laughs]

Harj Taggar 19:11 Yeah, you should tell us about the idea. [laughs]

Esha 19:13 Oh, our idea was to build an education platform for-college students in India to get jobs. And that was the idea we applied with, and I think you know this as well. Like, you know, when we came into the interview, your entire questions were like, "Oh, can you te- te- Varun, you have this experience here. Can you talk about like what other companies can you build from this experience?" And then we were, we, we were, like, so surprised. We were like, okay, we-

Varun 19:37 I mean, I, I rehearsed for this interview so much.

Esha 19:40 [laughs]

Varun 19:41 I mean, like, there are, like, 10 YC founders who interviewed me, and I know answers to all the questions, which is like, "Why this? What's the big tam and everything?" And you come to the interview-

Esha 19:49 [laughs]

Varun 19:49 ... and you ask her a completely different question-

Esha 19:51 [laughs]

Varun 19:51 ... which is like a pick a new idea.

Esha 19:52 [laughs]

Varun 19:52 And I was like, "What?" And it, it's, it's a, it's a-

Esha 19:56 Yeah. And then, and then our, our thought process was like, okay, YC likes people who are very gritty-

Varun 19:59 Yeah

Esha 19:59 ... so we should stick with the idea we came in with.

Varun 20:02 [laughs]

Esha 20:02 We're like, "No, Harsh." Like, "Yes, there are other ideas, but we'll do this." And then I think, like, one of the questions you asked was, "Oh, if we don't... If YC doesn't take you, are you gonna keep doing your company?" And then we said yes, but I think you knew [laughs] we wouldn't have done it.

Varun 20:16 No. [laughs]

Esha 20:16 Um, yeah. I think YC is one of the reasons, like, we even exist at this point in time.

Varun 20:20 I- is, uh, honestly, like kudos to you for just taking a bet on us. Uh, I don't thi- I mean, it, we got into this in some stage. I mean, like, we had, uh, uh, we both had, like, uh, high-frequency trading offers-

Esha 20:30 Yeah

Varun 20:30 ... w- paying us more than 600K out of college. And [laughs] I applied to a bunch of Ivy Leagues, and I got into [laughs] all of them. And at that point of time, YC was always this, like, North Star who, one company who, I mean, I thought that YC was one of those things which creates this massive generation of companies. And-

Harj Taggar 20:45 Yeah, and that was always something that stood out for me. I was always curious about this. I don't think I've asked you. But yeah, like, I mean, I feel like tr- turning down quant jobs if you're in, in America, in Europe and America is sort of like one level of risk to do a startup.

Esha 20:57 Yeah.

Harj Taggar 20:57 But by being, growing up in India and, like, turning down, like, super high-paying jobs [laughs] offers-

Esha 21:02 Yeah

Harj Taggar 21:02 ... like, must just be... How did you think about the risk of doing that, and were your friends and family supportive of it?

Esha 21:07 India is, like, generally, like, not a risk-taking. There's, like, very low risk.

Harj Taggar 21:11 And friends don't like their friends outgrowing them typically.

Esha 21:14 Yeah, true.

Harj Taggar 21:14 It's, it's always tricky.

Varun 21:15 My dad thought literally I was, like, crazy. I mean [laughs] ...

Esha 21:19 Yeah, I mean, maybe he was. He was right.

Varun 21:21 [laughs]

Esha 21:21 Probably, yeah.

Varun 21:21 I mean, like, uh, I mean, I cle- clearly remember I got, like, this highest paying job out of ever, out of all IITs. This high-frequency trading firm was paying me this much amount of money.

Harj Taggar 21:31 'Cause you guys had some of the highest ranking scores at IIT, right?

Varun 21:34 He, he, he, not me. [laughs]

Esha 21:37 [laughs]

Harj Taggar 21:38 Where did you rank at IIT?

Esha 21:39 I was, like, third in, in KGP Kharagpur. Um, but I mean, the reason we decided to do a startup, I think, like, for Varun, it was majorly around, like, passion, and he's like, "Okay, I just wanna do this. Like, even if I go to, like, you know, let's say these HFTs, I'm still gonna wanna do a company. I will do it, like, regardless of what I want to do." For me, it was more along the lines of, like, I was just logically convinced that, okay, just doing a company than working at, like, let's say these HFTs for three years is still a better outcome for me, even if I fail, uh, by going into YC. That was just, like, a

Esha 22:10 logical conclusion I came to. I was like, "Okay, sounds good. Let's actually give it a shot," and then, yeah.

Varun 22:14 And for me, to be honest, it was just doing the hardest thing possible. I mean, it translates to a lot of our company culture as well, just picking up the hardest problem and solve it. I thought genuinely, like, a lot of rich people and powerful people build [laughs] their own business, so I thought I wanted to do something around those lines, and I thought it was super hard. And yeah.

Esha 22:31 It's kinda surprising because this aspect of, okay, we wanna pick the hardest problems to solve, kind of translates so well, because we go into the sales calls, and then we talk to these, let's say, C-level people. A lot of the other competitors generally ask them, "Okay, what's the biggest chunk of support problems we can solve? How can we get you to 70%?" We don't do that. We go them and ask them, "What's the most complex issue you have? Let me go and solve that for you, and then you trust me to take over the rest of the entire support ticket." And I mean, this is my personal metric to build a product. I want to get every single one of my customers to 98% resolution, and not because I believe it's a very good, like, ROI. Yes, it is a good ROI, but I genuinely believe, like, you know, it's a better experience for the end customers to get, like, an AI which can solve a 98% of your issues. The biggest reasons humans still say, like, agent- a human, human, human, this is, like, classic behavior we see, is because there is no trust. There is no trust that AI can actually solve majority of the problems. But if you build an AI which can solve 98% of the problems, why would they not, like, you know, talk to it?

Harj Taggar 23:33 To me, it's just like the, the era to build a startup now is different, feels different post-ChatGPT-

Esha 23:39 Yeah

Harj Taggar 23:39 ... than before. And in particular, it's just there is a real return to being a, a deeply technical founder. Um, and you can just build your company a different way. By going in, I think it's, going in and just being like, give us your... Like, building a whole sales process around just give us your most complicated, like, most complex-

Esha 23:53 Yeah, and then I'm gonna show it to you that we can do it.

Varun 23:56 Yeah. [laughs]

Esha 23:56 Yeah.

Harj Taggar 23:56 Yeah, I remember. So I remember the interview. I remember you guys being very stubbornly attached to your [laughs] like, the college student idea. Um, talk us through, like, actually, um, going through YC and you, you, you even had, like, difficulties getting into the country.

Varun 24:09 Yeah.

Harj Taggar 24:09 Just tell us the whole journey.

Varun 24:11 No, uh, firstly, I was more curious on why did you take us in?

Harj Taggar 24:13 The cool thing we do is just invest in people. So it was very, very clear to me, even in the 10 minutes and from reading the application, the two of you are, like, extremely high IQ, um, uh, extremely determined. Um, but yeah, the fact that you were, you had mentioned that you were turning down the HFT offers and you were turning down... Actually, I was impressed by the fact that you were even willing to turn down, um, your research offer from Stanford. Um, so i- in general, all we do at YC is try and find people who are, like, unusual compared to their peers. And so the fact that you are, like, IIT, and yes, I'm not, I wasn't born in India, but my parents are from India, and I just know culturally it's so risk-averse, that I was like, wow, like, if they are willing to take this amount of risk, they are probably quite unusual people. And in the 10-minute interview, it was clear just that, not just that you were smart, but you just had a certain intensity about you. And yeah, one thing I've learned doing this for a while is the sort of... Yes, like, I didn't like the idea. It was probably very obvious.

Esha 25:01 [laughs]

Harj Taggar 25:01 Um, and it is a bit of a high-risk strategy to keep defending the idea anyway.

Esha 25:05 Yeah.

Harj Taggar 25:05 But I [laughs] ...

Esha 25:08 [laughs]

Harj Taggar 25:08 But I, I've, I've learned that if you're sort of, if you're just smart and intense enough, like, I think those people actually are always just sort of stubbornly attached to the idea. Um, this is actually, it reminded me a lot of the Amplitude interview. Amplitude is now a public company doing analytics, and this is, like, a decade ago. Spencer Skates, the founder of that company, also really, really high IQ, very intense person. He had an idea to build, like, a, um, voice assistant for you while you were driving, but this is, like-12. That-

Varun 25:35 12. [laughs]

Harj Taggar 25:35 Yeah, it wasn't really possible. Um, but he just defended it so intensely. I remember in the interview thinking, like anyone who's this smart and this intense, you just sort of have to invest in and work with them.

Varun 25:47 Mm.

Harj Taggar 25:47 And it took them, like a year and a half to find, like the Amplitude idea, and it took you guys like a little bit of time-

Varun 25:52 Yeah

Harj Taggar 25:52 ... as well. But I just... That, that intensity I think is quite rare. That's why, that's why we funded you.

Varun 25:56 Yeah. Uh, coming to the YC thing, uh, yeah. We thought it would be so easy to get into US, to be honest.

Harj Taggar 26:02 Yeah. [laughs]

Varun 26:03 That we'd go there. [laughs]

Esha 26:04 No, no, no. In, in the interview, I think when you gave us the offer, you were like, "Okay, when can you show up in, like, you know, in the US?" You were like, "Okay, we'll be there in a week."

Varun 26:13 [laughs]

Esha 26:13 And then, and then you were like, "Okay, are you sure you can be there in a week?" We were like, "Yes." And then actually, like, we could have gotten there... gotten here in a week. We did have like a visa interview scheduled, and then we went to the visa interview, like the... I think like the entire B1, B2 visa system is like, kind of like, you know, it needs a cleanup. But we went there, they asked us like, "Hey, are you married? Do you have kids?" We were like, "No." And they were like, "Okay, you're rejected." We were like, "What?"

Varun 26:36 [laughs]

Esha 26:36 And then we come back.

Varun 26:37 Your, your interview was at least like a little longer.

Esha 26:39 Yeah.

Varun 26:39 Mine was like just three questions: married, salary, "How much is your salary?" I was like, "I have my own company."

Varun 26:45 And I was trying to explain what YC is to that guy, and he was like, "No. It's done." For a minute I was just like, I did not even realize that people can reject you on a visa interview.

Esha 26:54 Oh, yeah.

Varun 26:54 I was like, "What's happening with my life?" It's just, um...

Harj Taggar 26:58 Yeah, I remember. So then, 'cause this was for the summer '23 batch.

Varun 27:00 Yeah.

Harj Taggar 27:00 And so then we basic- You had to do the batch remote, uh-

Varun 27:03 Yeah

Harj Taggar 27:03 ... essentially, right? So I didn't actually meet you in person, um, probably-

Esha 27:08 Until, like January 2024.

Harj Taggar 27:10 Yeah. So it was kind of... Which is getting close to, like, almost close to a year since, um, the, the first interview. Probably like nine to 10 months-

Esha 27:16 Yeah

Harj Taggar 27:16 ... after the interview.

Varun 27:17 The insane crazy thing was everything became harder, right? I mean, we don't know what our peers were working on-

Esha 27:22 Yeah

Varun 27:22 ... and it's, it's, uh, we... I mean, it's, it's our first time trying it out as well. Then we put it to fine-tuning, and we picked a really vertical where we can nail it.

Harj Taggar 27:30 That definitely... But on that note, it definitely felt to me like the you getting here, being to Silicon Valley, and being here at YC in person-

Varun 27:37 Yeah

Harj Taggar 27:37 ... definitely really helped accelerate the process-

Varun 27:39 For sure

Harj Taggar 27:39 ... of finding, like, the new idea, right?

Varun 27:40 A lot and a lot, to be honest. It's extremely helpful because, I mean, I remember back in I was India, this was, again, thanks to Garry Tan the reason the fundraise happened, right? I mean, I, I was asking my peers how much money they were, like, raising, and they're like, they're like 60 interviews scheduled for everybody.

Harj Taggar 27:54 Yeah.

Varun 27:54 I remember asking you, I was like, "Harj, I have only eight [laughs] eight investors. Am I doing anything wrong?"

Harj Taggar 27:59 This is for your Demo Day fundraise?

Varun 28:00 This is for Demo Day fundraise. And you were like, "No, no, you just go with it. Just go with the flow and everything." [laughs] And almost surprisingly, almost all of them end up investing. And Garry Tan told from our seed investor from, uh, Nexus, "You must invest in Giga." And they took a bet on us, and it's, uh... I mean, that really, that fundraise really helped us a lot because we're back in India and it's-

Esha 28:21 Yeah. I think, like, just having zero peers, and then, like, we both like, are like, "Okay, we just quit everything, and then now, now we are stuck in India. We want to be there." And then we don't even have, like... Especially with the education, YC really helped us because you connected us with a lot of people in education, education, and you were like, you were, like, pretty determined to get us off of that idea.

Harj Taggar 28:40 [laughs]

Esha 28:41 And then, and then we spoke to a few people. We were like, "Okay, yeah, it doesn't make sense. Let's pick something where we are good at." We pick, like, fine-tuning, and then the inference, that vertical. And then, yeah, it's majorly because of Garry, I think, our seed got done, pretty much. [laughs]

Harj Taggar 28:54 I remember... So, like, the timeline was basically you apply with the student idea, you pivoted into the fine-tuning as a service idea k- like-

Esha 29:01 In a, in a month.

Harj Taggar 29:01 In a month, yeah. So hitting almost the midpoint of the batch.

Esha 29:04 Yeah.

Harj Taggar 29:04 And then you raised at Demo Day on the fine-tuning as a service idea.

Esha 29:08 Yes.

Harj Taggar 29:08 Um, which actually I remember it got pretty decent traction-

Esha 29:11 Yes

Harj Taggar 29:11 ... like early on.

Esha 29:13 Well-

Harj Taggar 29:13 What exactly happened to that idea?

Esha 29:15 Okay, so for some context, we were basically training these open source models, fine-tuning them for specific use cases to get really good performance, outperform, like, let's say, Cl- Claude or GPT-4 or whatnot. And then back then, GPT-4 was the only real smart model, and it was, like, crazy expensive. If you remember the prices, it's like 32 thou- $32 for a million tokens or something. It had a good traction because people were optimizing for cost. And we were like, "Okay, it's good, it's good. People want to optimize for cost, and then we are gonna train these models, and you can use our inference." And then we had, like, decent revenue as well. And the biggest issue with that idea was it was like a broken funnel. People would come in, yes, they would be happy, like it's nothing wrong with the product. But then OpenAI or Anthropic, like, you know, they just release a new model which is like, let's say, as smart as GPT-4 but like, let's say, 10, 10 times cheaper. Which actually happened with, like, GPT-4.0.

Harj Taggar 30:05 Yeah.

Esha 30:05 It was like, I think, four times cheaper. And then we saw our customers, they were like, "Okay, we wanna move to OpenAI now because this new model can do these new other things." And our fine-tune model was trained for this use

Esha 30:15 case. We had to retrain it, and then it was just a lot of like, you know, continuous iteration we had to do to keep up. And it felt like we were kind of betting against the AI wave than betting into the AI wave. It's like, okay, if OpenAI stopped making better models, we would be doing really well. If OpenAI continuously made better models and cheaper models, we would be doing worse.

Harj Taggar 30:36 Yeah.

Esha 30:37 So that realization we got, I think, in February or-

Harj Taggar 30:40 Yeah

Esha 30:40 ... March 2024, we were like, "Okay, we don't wanna bet against AI. Let's just go into something where we bet on AI."

Harj Taggar 30:45 Okay. Um, and then how did you come up with the idea for Giga?

Varun 30:49 We were pivoting and trying out, like, a bunch of different ideas, and one of the... Some of the thing we're... At, at a point, we were trying to even be a software engineer. So [laughs]

Esha 30:59 Yeah.

Varun 30:59 We're, we're definitely, like, jumping on, like, a lot of ideas. Again, Nexus might introduce Zepto founders, and they told that customer, and that's when we got to know that, like, customer support is, like, one of the real problems, and they became one of the first customers.

Harj Taggar 31:11 Oh, so you got... You were introduced to Zepto, and you just went to talk to them in general about, "Hey, like, what are your-"

Varun 31:15 Yeah

Harj Taggar 31:16 ..."problems using AI?"

Varun 31:17 We, we, yeah, we tried, like, a little bit of hypothesis. I think there's one of our mutual friends who, again, another YC founder, who worked at Uber to automate their support, and they were saying, talking about support problems. I just made a hypothesis that Zepto might also [laughs] have similar problems.

Harj Taggar 31:31 Yeah.

Varun 31:31 It was a shot in the dark, and it kind of worked. [laughs]

Esha 31:35 Yeah, we, we spoke with them and then like, you know, they were like, "Okay, there's this huge event coming up in India. We need to hire a lot of people. We-- If we can help use AI to automate support, it would be massive help because we don't have to hire, like, what, 2,000 people or something." Um, and then we were like, "Okay, this is the first application problem." So we flew to India, we were in SF, flew to India, we went in-- we went and sat in their call centers for like, you know, days, and then observed how they support people do this work. And then the solution we built was, like, so custom, so random. It's like they didn't have APIs, so we built like a web automation solution on their CRM, which was very hard to scale, but we scaled it anyways. And then it was like a very custom solution for Zepto we built out.

Harj Taggar 32:18 And that sort of laid the foundation for the product as you have today?

Esha 32:22 Yeah.

Harj Taggar 32:22 Yes.

Varun 32:22 Yeah. At that point of time, we realized that we are not gonna do consulting again [laughs] .

Esha 32:25 Yes. Oh, yeah, no. We were like, "Okay, this is no way gonna scale."

Varun 32:29 Yeah.

Esha 32:29 And then that was where we got our determination, like, we have to build a product. There is no way we can scale this because... And then it was a very interesting product challenge as well. Like, we knew all the kind of support problems Zepto had, and we knew all the kind of support problems for a few other companies, and now the problem was: How do you build something the, in the most simplest way which can solve all of these different problems? Um, that was basically the product problem we had to solve.

Harj Taggar 32:53 Okay. So if we look ahead to the future a little bit, um, I mean, you guys are right at the cutting edge. Like, you're, you're operating at-- your agents operate at, like, probably, like, the largest scale of any agents in the customer support space. Um, where do you see this space evolving of AI agents in the enterprise in general, a-especially as the models themselves get more powerful?

Esha 33:10 This is the bet we are also making. Um, data and context is the, are the most important things, and AI gets better. The companies with the most context and data will be able to perform the best. Even if, like, OpenAI had a much smarter model, but it do- if it doesn't have the context, it cannot do much within an enterprise, and that's where we are heading as well. We start with customer support. We provide a lot of value to our customers. That sets us up to win a lot of these other operations within these enterprises. And the vision of our company is,

Esha 33:40 like, you know, this is, this is maybe, like, you know, a decade away potentially, but we wanna build a platform where the next trillion-dollar businesses are built on top of, and we wanna build the most efficient op-ops automation platform in the world. And we think we are, like, you know, very well-positioned because customer support is one thing people are looking to automate right now, and next things would be, like, you know, something much which involve more context. And since you already solve support, you have, like, all the context about potentially any kind of issue customers might face that will make you make better judgments in other verticals within the company. Um, so that's where we think, like, the market is heading towards, like, you know, potentially consolidation rather than, like, you know, point solutions for each of these AI verticals.

Varun 34:22 Since we have context from every single customer and what their past interactions and what are they looking-

Esha 34:26 Yes, and then, and then-

Varun 34:27 ... and even like both, like, very micro and macro level view of-

Esha 34:29 Yeah

Varun 34:29 ... what's, what all the customers are asking, it helps us solve other customers, other issues in the company very better.

Esha 34:35 Yes. It's like, I mean, fundamentally, every company is an optimization function on top of, like, value you give to your customers. And if you understand the value you give to your customers and the issues they are facing, you can write that function in a really good way.

Harj Taggar 34:46 For Giga itself, like, when you think about how big the company could be, are there, are there existing big companies that you look at that you're, you aspire to, either because of something they've done in how they built the product or the size of the company today?

Esha 34:58 I think we aspire parts of different companies. So Salesforce, I think we really aspire, like, the way they built the most general product, and I mean, obviously, like, you know, we don't like the design of Salesforce-

Harj Taggar 35:10 [laughs]

Esha 35:10 ... but obviously, like, they built the most general version of CRM pretty much. And OpenAI, I think they are always at the forefront innovating and things like that. So different parts of different companies are something we aspire, but I don't think we are looking to be like, "Oh, we wanna be the next OpenAI," or, "We wanna be the next Salesforce," or something like that.

Varun 35:28 Different parts of different companies. Like, I mean, Anthropic for some re- I think has a very-

Esha 35:32 Very good taste.

Varun 35:33 Yeah, taste and a lot of great talent compared to OpenAI right now who are going to them, yeah.

Harj Taggar 35:38 Varun and Esha, that was fantastic. Thanks so much for joining us. That's all we have time for today. Congratulations on the round, and I'm looking forward to seeing Giga continue to grow into a giant company.

Varun 35:46 Thanks, Harj.

Esha 35:47 Thanks, Harj. [outro music]