SFG 25: Anant Bhardwaj on automating data processes for the F100

‍

In this episode of the Startup Field Guide podcast, Sandhya Hegde chats with Anant Bhardwaj, CEO and founder of Instabase about the company's path to product-market fit. Instabase is an enterprise-focused application platform that allows businesses to automate document processing from unstructured data sources such as bank statements, pay stubs, and tax documents. They use a number of tools including generative AI to summarize and analyze data instantly. Over 50 large enterprises, mostly Fortune 100, use Instabase today.
‍‍‍

Be sure to check out more Startup Field Guide Podcast episodes on Spotify, Apple, and YouTube. Hosted by Unusual Ventures General Partner Sandhya Hegde (former EVP at Amplitude), the SFG podcast uncovers how the top unicorn founders of today really found product-market fit.

‍

If you are interested in learning more about the topics we discuss in this episode, please check out our resources on hiring your first sales leader, raising seed and series A funding, and hiring and build an early stage team.

‍

TL;DR

‍

• Anant’s journey as a founder: While Anant moved to the US to pursue a master’s degree at Stanford, he had always wanted to start a company. After he graduated from Stanford, he moved on to MIT where he met several influential mentors who encouraged him to pursue different research projects — one of these projects eventually became Instabase. 
• ‍The founding insight: Anant’s idea was to create an operating system layer for data to make applications portable and decouple them from underlying storage and schema information models.
• ‍Initial use cases: Instabase initially grew its user base by partnering with universities that needed to provide data science students with large data sets. Another version of the application involved converting log files into structured tables which found some traction, although early users expressed interest in making large sets of PDF files usable.
• ‍Iterating to product-market fit: After Instabase landed Standard Chartered, they began acquiring other major banks as customers. Initial conversations with these stakeholders helped the company understand how customers wanted to use the product and what problems needed to be solved. 
• ‍Growing the company: Instabase had a great deal of VC interest, but they raised money at too high a valuation before having a strong sales infrastructure in place. Building a sales team took time and money which curtailed the company’s growth for a while.
• ‍Instabase’s AI strategy: In partnership with Open AI, Instabase built its own models on top of Open AI’s language models so that no customer data is ever retained or used as a basis for training LLMs. Instabase has created an AI Hub which allows users to quickly analyze and automate content from files and databases. Instabase’s technical intervention allows LLMs to represent knowledge from uncompressed, loss-less datasets which minimizes the model’s propensity to hallucinate. 

‍

‍

Episode transcript

‍

Sandhya Hegde:

Welcome to the Startup Field Guide, where we learn from successful founders of unicorn startups how companies truly find product-market fit. I'm your host, Sandhya Hegde, and today, we'll be diving into the story of Instabase. Started in 2015, Instabase is a new entrant in the space of enterprise document processing and automation. Last valued at $2 billion, it uses several different AI techniques, NLP, OCR, computer vision, and now generative AI to extract unstructured data from documents, like pay stubs, tax forms, ID cards, and make them easily usable for massive organizations. They also help their customers in financial services, consumer goods, etc., to build applications that can automate workflows using this data. Over 50 large enterprises, mostly Fortune 100, use Instabase today, and we have with us the CEO and co-founder of the company, Anant Bhardwaj. 

‍

Anant, I want to go back to the origin story. I believe you moved to the Bay Area in 2010 from Pune in India for a Master's at Stanford in computer science. I'm really curious what your experience was like making that transition and how it has maybe informed what you are doing today.

‍

How Anant Bhardwaj became a founder

‍

Anant Bhardwaj:

I took a job in India after finishing my undergrad, and that was a time, I think, when a lot of cool companies were coming from Silicon Valley, so Facebook and Google, and if you read about the founders of several of those companies, they came from Stanford. Stanford was called the startup hub. I had no hope of getting admission there, but I still applied. I applied to a number of schools, including Stanford, for my Master's, with the hope that ... I always had this ambition of starting a company, and I thought that maybe going to Stanford is one way of doing it without knowing how it all is going to work out. I think in general, Stanford claims that they don't make mistakes. That year, they did, because my academics weren't good. They gave me admission. I'm surprised, because I got rejected from a number of other schools, their peer schools.

Sandhya Hegde:

Clearly, they knew something about you that was, to them, already worth the risk.

Anant Bhardwaj:

So, Stanford, I got admission, and I got admission into their Master's program, which is an MS CS. I moved to the US. This was my first time in the US, I'd never left India before, and I think the courses were okay. We studied the courses in Indian undergrad school, so the database class or the operating system class were not that difficult. In general, I knew those basic concepts, so when I took the classes, those were fine. In general, I think there was a different kind of focus here, more practical applications. For example, homework would not be, "Explain how the underlying architecture of something works," but more about, "Here is the application you want to build. What are the tradeoffs you need to consider? Build an application and prove that by running some code." I liked that aspect.

At the same time, I think this was a big culture shift in the sense that ... the way I saw things working in India, and I did spend a lot of time making sure that I'd do well in those classes, and retroactively, I don't think those classes mattered that much. I didn't do much at Stanford, really. I actually spent most of the time doing well in classes, and the rest of the time finding RAs so that I can pay money for the tuition because Stanford is a very, very expensive school. The reality is, I did not do a lot of research. I found a professor, Scott Klemmer, who was doing human-computer interaction. I liked the research a lot, and then I took a number of independent studies so that I could do some research, but generally, most of my time was spent just taking classes.

Sandhya Hegde:

Then you spent years at MIT. I'm curious, what was the switch for you? When did you decide you really want to start a company? What was the founding insight and exposure to entrepreneurship you had at that time?

Anant Bhardwaj:

When I graduated from Stanford, it felt like a last opportunity, because I wanted to start a company, and I spent last two years, literally, taking classes. I definitely had learned a few things, but not a lot compared to what I already knew when I was in India, although I made some really good friends and ... Stanford classes are much more intense, and I took some classes that I would not have been able to take in India, but in general, I felt this was a last opportunity. I came here to start a company, and I have a Master's degree and maybe a job offer from Google. I interned at Oracle and then at Google, and the choice was go back to one of these big companies, which I did not want to do. I really did not know what to do.

‍

Somebody, I think, had jokingly said, "If you really don't know what you want to do, go do a PhD, because that's a really failsafe environment to try different things and gives you some time to feel things out." That comment kind of echoed. I was like, if you don't know what you really want to do in your life, just do a PhD, and PhD doesn't kick you out in two years. You can spend a long period of time. Surprisingly, when I applied for a PhD, I got admissions from almost every single place I applied, so UC Berkeley, MIT, CMU, all the top schools, Harvard, Princeton ... wherever I applied. I like the Bay Area, so I wanted to go to UC Berkeley, actually, and I met a couple of professors, Ion Stoica, who did distributed systems, who started this great company called Databricks and Anyscale, and Joe Hellerstein. At that time, he had just started a company called Trifacta. I met them, and I wanted to actually move to UC Berkeley to do a ... I had a good collaboration with Joe Hellerstein, actually, because Jeff who was my advisor at Stanford, he used to collaborate with Joe Hellerstein. They both were starting their own companies, so they were spending less time on the research focus at their school and were more focused on some of the stuff that they were doing at Databricks and Trifacta respectively.

‍

I looked around, and then I felt maybe MIT could be a good option because I had already spent time at Stanford. Berkeley, I could not find a professor that I wanted to work with, even though I had admission. Other options were Carnegie Mellon and MIT. Those are all top-rated schools. I ended up moving to MIT and working with Sam Madden and David Karger. This is kind of interesting, because David Karger was doing human-computer interaction, very similar to what Scott Klemmer was doing, my advisor at Stanford. I also love systems. I love building things. Sam Madden basically ran the database group at MIT or the systems group at MIT. I felt like maybe this co-advising situation would work really well.

‍

I moved to MIT, and I told them that I had no interest in doing academia, because at the end, whenever ... Everybody who joins a Ph.D., they have ambitions to become a great professor. It was kind of unnatural. Why would you come to Ph.D. if you really don't like academia? I had made it very, very clear to my advisors that the reason why I'm here is I just don't know what I want to do, so this is a failsafe environment. I was honest with them, and they were totally fine. They said, "Yeah, just come here, do cool stuff, and once you've figured out what you want to do with your life, just do that."

‍

I went to MIT. I did a bunch of different, random, hacky projects, actually, rather than figuring out one research question and doing deep-down research. I built like five different research projects. One was called VoiceX, which was voice communication in a rural area ... These are completely different projects. The second was MailX, because social was becoming a big thing, which was how you can use group mailing in a way that allows social moderation and that kind of stuff based off all these mailing lists people use to spam, and people were learning that if you can apply some kind of social moderation the way Quora and those things work and have a different impact.

‍

Then I did another project called Confer, which is how people could meet each other in the business setting at conferences. That was widely used by a lot of conferences. We in fact onboarded several thousand researchers and students and people who used to join conferences. I think this is still working. MIT has kept it alive that project. The last one was this project called Datahub. The whole idea was how you can abstract this notion of databases and just create a single hub where you can connect any kind of data sources and build applications, and applications could be agnostic of underlying data models and all that and could be connected in real-time.

‍

These were the four projects, and I did all four of them in parallel. Datahub started getting a lot of traction in the sense that a lot of people wanted to build cool applications. MIT in general ... A lot of people from venture capital visit there and they look for cool projects. There was this investor called Alex Taussig at Highland Capital at that point, Now he's at Lightspeed. He told me, "Hey, why don't you drop out? This seems like a very cool project." I asked him if he would fund me if I finally dropped out, and he said, "Yeah. Drop out. This seems like a very good project. You will easily get funding." One lesson that I learned is don't trust things investors say in general. I literally dropped out and went to him, "Hey, I dropped out. Do you want to give me money?" He said, "I have to talk to my partners." It wasn't that good, although I was still able to raise money and all that, but hopefully, that tells you the story of how this idea of starting a company came from. So it just became a project that really started getting some traction.

‍

Then I went to my advisor and told him, "Hey, I'm leaving." I think there is one sentence that Sam Madden said, my advisor, and it was such a profound sentence that is still a reminder of ... whenever I feel confused. He said, "You can stay at MIT, finish your Ph.D., write some research paper, get a great faculty job, I'll be proud, or you can leave today, go build a company, make it very successful, I'll be equally proud. Just don't do something shit." Then he signed my paperwork. That was the story of leaving MIT.

‍

How Instabase found early investors

‍

Sandhya Hegde:

That's a really good quote. That's a great compass to have in your pocket. I'm curious, how do you remember your beliefs and maybe assumptions about being a founder at the time? You still had spent a lot of time in research, on campuses ... the really great ones, but hadn't worked in a startup, you didn't have a CEO founder role model. What did you remember at the time? What was the assumption going in about how this would work? Did you look for co-founders? How did you think about the act of starting a company and validating your unique idea?

Anant Bhardwaj:

I think my answers would be something which is not going to be very helpful to a lot of people. I really didn't know. The reality was that I just felt it was a cool project. That's the only thing I knew. I really did not know who would use it, why they would use it, and where I could make money from. When I talked to these VCs ... Matei Zaharia, who is the CTO of Databricks, had just become faculty at MIT, and I asked him, "Matei, it seems like Databricks just got funded by NEA and Andreessen. Can you make an intro?" Matei made the intro with Peter Levine from Andreessen and Pete Sonsini, from NEA, and they both wanted to do a deal. At that time, Andreessen had this philosophy, and I think they still have ... They will not invest in a company if that has a conflict of interest with one of their existing portfolio. He wanted to do the deal, but one of their existing companies said they have a conflict with Instabase, so Andreessen pulled the offer with a lot of humility and respect because of their policy. Pete Sonsini connected me with Jerry Chen from Greylock, and Pete and Jerry together did the seed round.

‍

We really didn't know at that time. The pitch was that we wanted to create this operating system for data where people can build applications, and what applications, who will use it ... I really have no idea how they end up deciding to invest. Also, this was kind of a funny and bizarre story. At university, you don't make much money, so I dropped out with very little money, like $4,500 of savings, and I got sued for some old car accident that I had, and I had to settle that lawsuit for 4,500, so I literally had zero dollars in my bank account. I stayed at a friend's house and would take his bike and take the Caltrain to meet the investors in Sandhill because that's where the old investors used to be during that time. The city was still a new phenomenon. Most investors' headquarters used to be on Sandhill.

‍

Pete and I had a good meeting, and he said, "Most important meeting that is for you is going to be Forest Baskett, because he's our technical person, and he is going to evaluate Instabase." I'm like, "Sure." The meeting was supposed to be from 4:00 to 5:00 PM on ... I think Wednesday or Thursday. I took Caltrain, got out at 3:00, and realized that the tire of the bike was punctured. This is a friend's bike, so you can't leave it anywhere. I basically decided to walk the bike and really ... I reached 45 minutes late for the meeting. The meeting was supposed to be from 4:00 to 5:00. I didn't have money to ... My only options were Uber or Lyft or something, but I might not have had that account set up... I reached the meeting at 4:45, so 45 minutes, but Forest was still working in the office. I asked him, "Do you have another hour so we can actually chat?" He said, "No. I have another meeting starting in 12 minutes." I was very sweaty, but we chatted, and we discussed only for seven to eight minutes, and he still decided to move forward, which was super nice of him.

‍

Basically, NEA and Greylock did the deal with literally no idea who will use it and how they will use it.

‍

How Anant came up with the technical insight for Instabase

‍

Sandhya Hegde:

What was your technical insight at that time? What was the answer to how is this going to work that you are uniquely capable of building a solution for, even if we don't know exactly who will be the customer or what exactly will be the product? I'm curious, what was it you felt like you understood at the time that maybe the big existing incumbent companies in this space were just not approaching the right way?

Anant Bhardwaj:

I don't know. I think sometimes, you just get lucky, and I think it was in that case. The pitch was literally the way Windows or Mac operating systems abstracted all of the complexity of the underlying details of how microprocessors work and other things and just give you an application layer, and people can build applications, and it gives you the simplicity. That's, I think, what we need in data. We don't need 2,000 different tools. We want to be that operating system layer where people can build this stuff, which will make applications completely portable. Let's say you build a really good MapReduce program. If you want to apply it somewhere else, you have to really replicate all of the underlying architecture and the schema and information models and everything. We felt this is a very tightly coupled system, and we need this layer called the operating system layer, which is going to decouple the application from the underlying storage and schema and information models and those kinds of things.

‍

I don't know if there is any deep insight in that. Anybody understands that physical and logical independence of the storage layer from the application is a good thing. Maybe, I think, sometimes, if you studied at MIT and you say something, people find it more profound than it really is. I think that was the case. Maybe investors felt that whatever I was talking about is more profound than it really was, but the reality is, I don't think I had understood anything more than what other people had understood. I just think that initially, all people take a bet on a person. I was very clear: I don't have any customers. I really don't know. Here is what I've built. I believe a lot of people want to use it. They can build applications. They love building applications for it. I have not written a single line of code. Everything is an MIT research project. We'll go and do it.

‍

I think people believe in that. Maybe I got lucky. I know a lot of my friends who were much smarter than me and had built out a lot more sometimes get scrutinized a lot more than I got scrutinized during that time. The seed round was super easy. There is a blog written by NEA, I think, which is ... They asked, "What's this guy doing? Nobody understands. Somebody said he wants to build this ... the operating system for the Internet where you can just connect any data and anybody can build applications and you port around." The reality is I don't think there was anything profound there, it's just some idea that looked very grand.

‍

The early use cases for Instabase

‍

Sandhya Hegde:

Okay, so you have a seed round for your grand idea, now you can at least not be broke anymore, you can start building a team, building a product. What were the first six months like?

Anant Bhardwaj:

It wasn't as luxurious as you think. It doesn't really pay to raise money. so I raised money, and then ... I was on a student visa, which basically means you have to get some work visa to be able to do the work. It gives you, I think, three months before you figure out the situation. There is something called OPT, Optional Practical Training or something, but you get that only if you finish your degree. You don't get that if you don't finish your degree, and I did not finish my degree. That was the catch. I applied for H1B and it got rejected, so I got a notice to leave the country. Now you have all this money and you cannot work. The reason why it got rejected was that you need this employer-employee relationship to work. You cannot employ yourself. Anybody can start a company and then can employ himself.

‍

At that time, MIT came to my rescue. My advisor admitted me back while we had the legal team figure out the situation. It took about two years to figure out that situation. We had to go through a number of back-and-forth legal procedures. Eventually, I got H1B, and my advisor basically got me admitted back into MIT and kept me on a student visa while I was working for the company for the next, I think, year or two. So MIT basically saved me from getting deported ... I don't know if it's legal or not, but they actually went out of their way to help me.

‍

There was only one person, just me. There was no co-founder. Investors were like, "This is four months, the guy has raised money, doesn't have any person ... What is he doing?" I took time. I take a lot of time. I started meeting a bunch of people, and then I found a person from Google who I really liked. We worked for two to three days, and we hired him literally around the end of November of 2015 as the first employee. Our first set of use cases were basically ... I don't know if you have seen the Jupyter notebook that people use for data science. It was literally a notebook app making this work on our, basically, operating system.

‍

That particular app, people liked, and I'll tell you why people liked it. One of the issues that I think all the universities had during their time was if you have to give homework on very large data sets. Nobody can fit those stuff in a laptop, and there was no good way of running this ... You could run the notebook locally, but if you can't fit the data set, then you are in trouble. How do you give students basically a way to do this where they can still do this homework but on a large data set? They started putting this massive ... not super-massive, but hundreds of gigabytes of datasets, like an entire wiki dump, and asking students to basically do this. They were getting free computing. It was really losing money. We gave basically a lot of students at Stanford and MIT and the University of Chicago and Columbia doing this homework still.

‍

The dean of Stanford, Jennifer Widom, at that time, she decided to basically teach worldwide some of the classes at Stanford. And she used Instabase for, I think, several years, two or three years, but we were losing money. The students don't want to pay money, and this free notebook, free computing, free storage, and free database we were giving wasn't accruing any benefit for the company, but we got some good number of users, we built something that people were using, so we were feeling good. It was about ... almost one-and-a-half years or so, and we figured out that it's hard to make money from this app. This app is not the app that is going to make us money, even though people love to use it.

‍

We built some other apps. One of the apps was called Refiner which basically can take your log files and automatically convert them to structured tables that you can do analysis on. We called that app Refiner. We used to just go and do the cool demo ... Whoever comes to Instabase is literally going, "So, cool demo." so one guy who was from this company called Zenefits, they do HR and payroll, he was seeing the demo and he said, "This log file is cool, but our problem is we get a bunch of these PDFs from health insurance providers so that our people can pick the best health insurance plan, and frankly, we have to manually put those things so people can compare, because otherwise ... Nobody wants to read a 100-page PDF. Can you do that?" We were like, "We don't deal with PDFs. We are an operating system." We can add an app that reads PDF and converts it to text and refines that — We added that. Zenefits went through a leadership change during that time, so we never closed the deal, and there was a CEO change.

‍

Then we were doing the same demo with a friend who used to work at Lending Club at that time. "We have a similar problem. When people apply for a loan, they submit pay stubs and bank statements and 1040s and drivers licenses and passports. How can we do that?" We were like, "We don't do OCR and those kinds of things, but we're an operating system." We just assembled some open-source stuff, put something up, and built a cool demo. Lending Club was going through some CEO change during that time. We were like, "Whenever we get close to a deal, they do a CEO change. That's not good."

‍

At that time, Martin Casado, who'd just joined Andreessen, he just visited, and he saw a demo of the product. Somehow, he was super impressed, and so he wants to do a Series A at a pretty high price. Sounds great. He wanted a Series A, then somebody suggested, "Hey, you already have an offer, maybe you should talk to some more VCs?" And then Peter Fenton from Benchmark, he showed a lot of interest, and a bunch of other good funds, Founders Fund and others, so it became competitive, and then we got a pretty good deal. We got a Series A. A couple of startups were using us at $2,000 a month or something they were paying us.

‍

At that time during that process ... Andreessen used to do EBC, and there is a company called Standard Chartered visiting them. Martin told them, "You should use Instabase," and then they called me in Singapore, we did some proof of concept, and they were like, "We want to do a deal." The product that they were trying to use earlier was IBM Watson, and the license fee that they were paying was something like $1.2 million a year. I'm like, "That stuff doesn't work, and you were willing to pay $1.2 million. This stuff works, so you could be paying more." We basically gave them a price of $1.5 million a year, and they actually said yes. Basically, now we have this price of $1.5 million, a bunch of these startups are paying $2,000 ... I'm like, "Look, the price has gone up to a million dollars," so those startups left us. They said, "We can't afford it."

‍

In fact, in Series A note, Frank Chen from Andreessen wrote a note which says, "This company that Martin wants to do had three customers before Series A. They have zero after Series A." Basically, this happened during the due diligence process, after the termsheet before the ... but Martin still went ahead and did the deal.

‍

That's basically what led to the product-market fit. In summary, the product-market fit now ... We got other banks with similar problems, and we started talking to them, but the reality is, and I write that in detail in my Series B note to investors ... A lot of people think that we were smart and we came up with the idea. That wasn't true. The reality was we really didn't know what we were doing for a good period of time. However, I think there is one thing which was true, and that is we always believed that we don't really need to be good at predicting the future as long as we remain curious, fearlessly experimental, and engage with the world with the right mindset. If we do that, the market will pull you towards the right product. It's just that you have to recognize where it's pulling you. If you can recognize it quickly, you can find the product-market fit.

‍

How Instabase found product-market fit

‍

Sandhya Hegde:

The fact that someone like Standard Chartered was willing to even work with a startup that didn't have any enterprise customers is a big signal for you that they really, really care about the problem.

Anant Bhardwaj:

Yeah. I think the product-market fit is just establishing if the problem is real. I don't think you need to really have the best product at that point in time. Just engage with the world to really establish that the problem is real, and I think you can build a good product. You have smart people. We always attracted smart people in hiring. That's how we found it. I don't know if we still found it. Still, every sale is a grind, that's the reality, but then we of course got ... almost four out of top five banks are our customers now, and globally, also, if you look at top 10 banks in the world, eight of them are our customers. Generally, I think we got on the phone and we got all of these people to start using the product, but it wasn't ... We didn't have product-market fit. We really didn't come up with it. It's just that over the period, after working with a number of people, we were able to understand what problem people are willing to pay money for, and we just doubled down on those. Not a profound answer. It was, I think, brute force.

Sandhya Hegde:

Now you have made this staggering transition from having a free app marketplace that students and small startups are using directly to Fortune 50, some of the most traditional, conservative, large institutions in the world ... How did you approach even learning what is the right go-to-market motion? How do you shape a sales process? Who's the right buyer? How did you even forecast how things are going to shape up for you after 2017, when you're focused on really large companies? Could you share more about that?

Anant Bhardwaj:

Yeah. The reality is that we made a ton of mistakes, so we can tell you all the things not to do. I think that we know more than all the things that you should do.

In 2017, we had no customers when we did the Andreessen one, then we started with Andreessen the Standard Chartered, and we closed the deal, I think, in early 2018. It took about six to eight months to land a deal. After that, we got some really good customers, some names that are the beacon of financial industries. I will not say the names, but just think of one of the top two or three banks. We got pretty quick success. After Standard Chartered, we got three more banks within the same year.

Sandhya Hegde:

Who were your champions in these banks?

Anant Bhardwaj:

Andreessen would make the intro, we'll just go, do the demo, and then ... I was the only person. The company was literally a four-person company when Andreessen invested in the Series A. In 2018, we were an eight, nine-person company, really, and ended with something like 20 ... We wanted some people to help with customer support and customer management and those kind of things. Literally, Andreessen used to do a lot of early seeds, and investors that they wanted to show that they can provide value, so they would make a ton of intro, and they will go and talk to a bunch of those people. Literally, early on, Brad Kern, who runs the Andreessen agency in New York, he will go with me to the customer. He was my salesperson for some period of time.

‍

That basically gave us an illusion, actually, in the sense that this company was growing from 6,000 revenue to four-and-a-half million in a single year ... That's a very staggering growth. 4.8 or 5 million dollar worth of revenue, I think, early 2019. Sarah Cannon from Index came and ... I didn't want to raise the money, but if somebody pays a billion dollars, maybe we can talk, and Index said, "Yeah, let's chat." I don't think it was a mistake, because it helped. I have a newfound respect for Sarah Cannon, because she took that bet on us, and that has shaped the company. I think that actually, we raised the money of Series B in 2019. I think we announced in July 2019 was around $105 million at a billion-dollar valuation. We were about 20-person company at that point, no salesperson, just all engineers, and there were four or five people in ... we call them the leverage team, which is like ... can talk to the customers and all that.

‍

Then we hit a roadblock, because we didn't know how to scale sales. Typically, when investors put so much money, you have this pressure to build a sales team quickly, so ... The typical suggestion that somebody would give you is go and hire a CRO and they will go and build a team. In general, I think you have to ask the question, "Is the company ready for a CRO?" We hired a CRO, and ... CRO typically also want to show that they are great. They're like, "We gave the forecast for the year 2020 as $20 million." The person joins around November and December 2019, and around March and April, I'm like, "This person has not hired anyone. How are we going to hit 20 million?"  If you take six months of ramp time ... It's just not going to work. You know the process, the capacity-based model.

‍

It became very clear that we have made the wrong hire, so we decided to part ways, and then, basically, we started building smaller VP-level hires. We hired Luke Rogers from AppDynamics around, I think, July 2020. That year was a brutal year. Literally, we went from 4.5 or 5 to 7 million. It's expected, but now all of my focus was hiring and building team, and I could not do sales. Earlier, I was the one who was doing the sales, so that year was basically a completely lost year. Of course, after we rebuilt the team in 2021, we went from 7 to 21, and then we doubled again after that, but we made the mistake that ... The key point, I think, is that we should not have raised at that high price, given that we had not built a team, but as a CEO, if somebody's given you such a high price, it's just very lucrative and it's very tempting to go and do it. In our case, it turned out to be all good, and I think, in fact, I'm happy that we raised that money, because it allowed us to really go and make mistakes and build that-

Sandhya Hegde:

You had a lot of buffer and runway during a period where the market was very constrained in terms of fundraising.

Anant Bhardwaj:

Yeah, absolutely. Basically, during COVID, the market had ... We had seven years of runway. The team started very small. We were only a 30-person company. That definitely helped. We were still making five million. In fact, we were cashflow positive overall until mid 2020. The key point is we were still making $5 million a year, and we're only a 15-person company, so we were not spending a lot, but the key point is we also did not grow. We grow from five to seven. After, of course, we tripled and then doubled, but I think we did not know, and we made a ton of mistakes, and eventually, I think we learned the lesson. Sales is... don't take it for granted. You have to actually build it and build it with a lot of logic, with a lot of reasoning, with a lot of measurement. If you don't do that, you will burn a lot of money. We have basically made so many mistakes in sales. The amount of wastage that we have done in terms of the cash to build a sales team is pretty mindblowing.

‍

How Instabase created its AI Hub

‍

Sandhya Hegde:

That's why you need the venture capital funding, otherwise ... I'm curious. You have always looked at multiple ML approaches. That's always been your secret sauce as you're not stuck with, "We do one thing well," it's more, "We enable you to choose whatever approach you need for your problem and workflow." When did you first start looking at incorporating LLMs as one of these approaches, and what was the reaction from your very traditional customer base? Were they resistant to the idea, excited about it? Could you help us understand what the enterprise adoption at that level of this new technology looked like?

Anant Bhardwaj:

I think we had the operating system, so that was the advantage. We had built our own distributed operating system. That basically means we could add and remove things pretty easily. OS can add and remove things without changing this basic idea that we suggested earlier of portable applications, because all the changes you can make in the operating system turned out to be very, very helpful to us. In fact, earlier, we were against ML, because ML did not work reliably for these unstructured data problems. In fact, in a customer meeting, somebody said, "Do you use ML?" I'm like, "Do you need ML or a solution to the problem?" They'll go silent in the room.

‍

Things changed around late 2018 when the Transformer paper came, Attention Is All You Need. I think there were some early language models, not LLM, but smaller language models ... GPT-1 and GPT-2, some of those models came. Transformer architecture really changed how do we approach these language problems? At that time, there was a board meeting, and I sent a note to the board. That was wildly contested at that time. We said that we have something like that ... "At Instabase, we have determined that deep learning is the future of language understanding. We are going to invest fully in that. The question is not about whether this approach will work or not. We have determined that we will invest everything to make this approach work."

‍

We wrote our own language model. The problem with the sequence-based language models was they basically represent language as a sequence of tokens and produce tokens, but it does not work well on documents. Let's say you're looking at a pay stub, then the rows and columns become very important. We came up with this idea of how you can encode the layout, the X and Y coordinates in those language models. We did a lot of good work in layout-aware language models where you basically also have the spatial representation, so you encode geometry in the language itself. And that document is ... basically, geometry becomes very important, because if you don't understand geometry, this is a jumbled set of words, and they might not mean anything. Documents use vision and language both together. That's why language becomes very important.

‍

We basically wrote that and we moved to a language model called InstaLM and the problem is in order to ... That was a foundation model that could be fine-tuned on any given problem by just taking 100 samples or something. It could run on CPU, but in order to fine-tune it, you needed a GPU. One of the big struggles that we had was many of the customers don't have GPUs, they didn't know how to operate GPUs. Our core proposition was you can run anywhere, on-prem ... because it's an operating system. As long as we can run containers and Kubernetes, we should be all good.

‍

We struggled in some places, but we had still ... sometimes, you just make a call that ... Look, we are going to go in that direction. We will invest in the rising sun, not the setting sun. Certain customers that were not ready, we were like, "Move to the cloud. Just do training on the cloud. You can still run those things on-prem." That worked out okay in most cases. Some customers still struggled, but in that case, we will basically try to get some synthetic data outside, train a model using GPUs on the cloud, and then run that model on-prem. That worked reasonably well. It's just that something that will take an hour to train on GPU, it will take a few days on CPUs. In the worst case, you can still run on CPU, it's just that the order of costs basically goes significantly up.

‍

This worked great until early 2022. That's when the large language models ... People basically started building language models that were much, much larger. A very good friend of mine is Ilya, who co-founded OpenAI, perhaps one of the most influential figures in everything that you have seen in the LM space, GPT-3, ChatGPT and all that. We were just chatting, and I'm like, "Let me experiment with some of the stuff that he is doing." We started encoding layout on large language models, and we saw great results without fine-tuning. I said, "Customers aren't sending data outside. This is going to be a very problematic endeavor." I talked to him, and he said, "Let's talk about some partnership, and we can figure out how to make this work." That's when he connected me with the partnership team of OpenAI, and we figured out how to get access to their models so that no data is ever retained, they cannot learn from it, and all of the legal stuff that we needed to get done.

‍

We were doing the experiments most of 2022, and once the results came good, we decided to now take a very clear stand that ... and it's a controversial stand, because I think there is just a lot of cool stuff that's happening in the open-source community that a few large language models will dominate rather than having multiple fine-tuned stuff. In fact, we wrote that fine-tuning is not going to be the future, and companies that will invest in a community of a large number of small models may not be the future. That's why we basically fully invested in AI Hub. AI Hub basically has some controversial bets, and I hope those turn out to be true, but we encourage those things. The key idea here is now given AI how it's publicly available, anybody can go and play with it, now we can open it for third-party developers and all that.

‍

I think you've talked about this question of ... The core question that basically came, which we addressed, was about data privacy and security, but then there was another question of trust and hallucination and planning with LLMs, and how do you make that work? We did a lot of work, and because of IP constraints, I can't give you everything, but I'll give you the high-level. We did a ton of work on fixing hallucination, and I'll give you the rationale for how we approached it.

‍

A lot of people believe that hallucination is because these large language models sometimes don't know how to reason things well. Our experiments suggest that that is not the reason. Just think of what these large language models are. They encode two things. One is they learn how to reason about the world. Second, they compress all of the knowledge that exists in the world into that little network. That's how they're able to give you the answer. I believe that this compression is very lossy. That means when you compress all the knowledge in the world, you lose many of the relationships. It's like Google right, it compresses all of the knowledge in the world, but you only have which document contains what words in what order. You lose a lot of other information. This is very ... Compression in many cases are lossy. It takes a lot of work to find lossless compressors.

‍

What if we don't compress things? It's not possible for the Internet, but let's say I'm working on a folder of documents, or let's say somebody applied for a loan and they submitted 20 documents. What if, on the fly, you can create embedding and the representation of that knowledge, which is uncompressed, now that is just created, and then reason on that uncompressed representation? You're bounding the knowledge to be only that, and you will not get the hallucination problem.

‍

I was surprised the kind of use cases that people have figured out. I was in India, and we were talking to some press, and there was this guy who does this mental health counselling. He said he looked at ChatGPT, he was very excited, and then he wanted to expose some of those things as a conversational app on phone for his customers where they can talk about mental health and they can get the right advice, but he said, "I have no control over what answers ChatGPT is going to give. That's not what I want. I've written these 15 books and last 10 years of notes. I want everything to come from that and nothing outside that." At Instabase, you can do that. Now you basically create the representation of those embeddings and the knowledge that ... It's just based on that dataset. If you basically force that the reasoning cannot go beyond that bounded representation, you will not hallucinate. How do you do that? It's a hard problem, so I'll not get into the technical detail-

Sandhya Hegde:

Yeah, makes sense.

Anant Bhardwaj:

I think the hallucination problem would continue to adjust when you are looking at the knowledge ... basically, the entire Internet, but there is a way to address hallucination when you bound the domain to something a lot more controlled.

Sandhya Hegde:

Awesome. This was such a great journey of product evolution and maturation. I'm curious what your advice would be for early-stage founders just getting started in AI right now. The vast majority of new companies being started are exploring how to leverage this big platform shift and all the new tech that has come out in the last couple years, so I'm curious, what would be your advice to someone new getting started right now?

Anant Bhardwaj:

My advice has always been the same. Don't spend too much time predicting the future. Just do what you feel is ... What you like exploring, just go and do it, and we'll figure out something. In general, I think just be curious, be experimental, and engage with the world where you can understand and recognize what problems are real and what feedback you are getting and quickly be able to react to it. This is just a meta-answer rather than an actual answer, but I think that applies much more than ... Nobody really knows what ... If somebody really knew how to create a startup, they could create this factory. The reality actually is that it's just ... Engage with the world with the point of view that you don't know everything, recognize the feedback to really establish which problems are real, and double down on that. I think that works in most cases.

Sandhya Hegde:

Thank you so much, Anant, for joining us today and all your candor. I think it will be an incredibly helpful story for founders to listen to, because so many of them are exactly in the shoes you were in in 2015 and trying to figure out, "Should I keep going? Should I stop, and which direction?" I think this is just a really amazing story. Thank you so much for sharing it.

‍