Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
How to Build Real-World AI Products That Scale | Pawan & Paramdeep (Shorthills AI)
236 Plays1 year ago
"The real power would lie if we crack the technology piece very clearly."
Pawan Prabhat and Paramdeep Singh emphasize the importance of robust technology in their ventures. In this episode, they share their journey of building and scaling tech-driven companies, highlighting the critical role of innovation and adaptability in solving complex business problems.
Pawan Prabhat is the Co-founder & President at Shorthills AI, a company focused on building enterprise applications on LLMs. He is a serial entrepreneur; his previous EdTech venture, EduPristine, grew at 80+% YoY and was acquired by Adtalem (NYSE: ATGE).
Paramdeep Singh is the President & Co-founder at Shorthills AI. He brings extensive experience in technology, operations, and business modeling. Previously, he co-founded EduPristine, which achieved market leadership in the Ed-Tech space.
Key Insights from the Conversation:
- The importance of having a long-term vision.
- How EdTech was not the flavor of the season when they started EduPristine.
- The intricacies of building and scaling AI-driven solutions.
- Why LLMs are not just hype.
- The future of generative AI in solving enterprise problems.
Chapters:
00:00:00 - Introduction to Serial Founders and Their Journey
00:01:12 - From Hardcore Coders to Business Leaders
00:05:36 - Choosing Between Online and Offline Education
00:08:39 - Finding Product-Market Fit
00:13:34 - Building a Profitable Business vs. Burning Cash for Scale
00:18:45 - Cracking the Global Market
00:28:48 - Starting Shorthills AI: Solving Tough Problems with AI
00:39:28 - The Hardest Problems in Building an AI-Powered Review Platform
00:51:57 - Evolution of Large Language Models (LLMs)
01:10:11 - Leveraging Generative AI for Enterprise Solutions
01:13:17 - The Nuances of Prompt Engineering
01:28:50 - The Long-Term Vision for Shorthills AI
Hashtags:
#GenerativeAI #LLMs #ArtificialIntelligence #MachineLearning #StartupIndia #FounderThesis #Entrepreneurship #TechStartups #AIApplications #DeepLearning #Innovation #BusinessStrategy #Podcast #Technology
Recommended
Transcript
Journey from Investment Banking to AI Startup
00:00:00
Speaker
Hi, this is Bhawan and I'm one of the co-founders of ShotlStick. Hi, this is Parandip. I'm one of the co-founders of ShotlStick. It's great to be with Akshay and Bhawan to discuss about generative AI today.
00:00:24
Speaker
What is a large language model? What is the meaning of the term GPT in shared GPT? If you are overwhelmed by the sudden barrage of AI-related jargons and want to really understand the hype behind AI, then this episode is a must-listen. In this episode of the Founder Thesis Podcast, your hosts Akshay Dutt interviews Pawan Prabhat and Paramdeep Singh
00:00:45
Speaker
The founders of the AI startup Shortheels AI. Pawan and Paramdeep are investment bankers turned founders and they built their first startup way back in 2008, before the startup boom really hit India. They built up a profitable edtech startup much before any of the current crop of edtech unicorns and were eventually acquired by a global edtech giant. In their second innings, they decided to focus on AI and build AI-powered products for themselves and their clients.
Educational Background and Early Career
00:01:12
Speaker
Stay tuned for this masterclass on all things AI and subscribe to the founder thesis podcast and any audio streaming app to learn more about the fast-changing technology landscape. Both of you are serial founders. I'm curious to understand the history of how you met and started your first venture. Let's talk about the first venture for a while and then we'll come to Shortels.
00:01:45
Speaker
So we go a long way in terms of friendship. So we, in fact, both of us are classmates from I am Indore in five. And then that's where we studied together. We published a couple of research papers together. And then both of us worked together at Santa Chara Capital Markets. And then we started our first company called EduChristine in 2000.
00:02:13
Speaker
and we've done that together for 10 years, in 2018. So, both of you are techies. Prior to I'm indoors, you both graduated from the IITs. Correct. That's right. So, both are techies. I worked with a startup called First Train. And I worked with a company called Geometric Software. This was a company which used to make product in the ad game space.
00:02:40
Speaker
And so, yeah, two years. So before an MBA, both of us were, you know, hardened Java, whatever program was, you know, this is what we were doing.
00:02:50
Speaker
Okay, okay. You understood how to code, basically. You were hands-on in that.
Founding and Growth of EduChristine
00:02:55
Speaker
What led to the idea of Edubristi? And, you know, this is like in an era where EdTech was not the flavor of the season when you started this out. Just help me understand what you saw as an opportunity and how that got off the ground.
00:03:14
Speaker
So when we started, both of us were working as investment bankers with Standard Chartered. And both of us had the experience in finance and quantitative analysis. So I think that's where our love for data and finance came. And once we had worked for a few years, we thought that it made sense to utilize this knowledge and help more professionals gain the same knowledge. So we had studied at IMG.
00:03:40
Speaker
What we found is that most of the courses tended to be very theoretical in Asia. And there was a large gap between what practically the industry wanted and what the colleges provided. So that's where we started. So we again started with professional certification called FRM, this Financial Risk Management. And gradually we've built our skills around other examinations like CFA. And then we sort of brought in our own certification like
00:04:04
Speaker
financial modeling in Excel. So all of those courses tended to be very sort of professional oriented, very job oriented. That means people who had the relevant knowledge from college, they could augment their knowledge with these courses and then become more employable. How did you get the conviction to quit your jobs?
00:04:26
Speaker
And did you want to do it online right from day one or did you want to send the course and online was just a distribution channel? So I think to answer the first question again, both of us will have different views on this. So before starting the company, we were in standard chartered capital markets and both of us were book running lead managers.
Inspiration to Become Entrepreneurs
00:04:49
Speaker
Our job was to take companies public.
00:04:52
Speaker
And day in and day out, we were meeting founders of companies which were about to go public. And since we were also taking companies public, we were meeting so many founders. Somewhere we thought, if not today, then when should we do that? Because we were meeting people who had taken a bet at least 10 years ago or 20 years ago. So I think that was one of the key motivations, at least for me.
00:05:15
Speaker
Fine, take the plunge. Otherwise, this will never end. That was the idea. How do you eventually make that call? So, that is one and I think, just to add to, you know, Paramdeep's thoughts from the previous point, I was always somewhere fascinated with the NIAID, you know, to that extent.
00:05:35
Speaker
I know you had already interviewed Mr. Rajendra Pavari on the podcast, but my analogy was always that there is an NIIT for tech guys, which offers good courses. It was presented across the country.
00:05:51
Speaker
There were institutions which were training people for MVA, like, I don't know, career or something. But there was nothing for finance guys, you know, professional services, certifications of finance. Personally, I thought there was a huge gap out there because there was nothing, you know, equivalent to NID, you had Aptech, you know, I think that was where my thoughts came in, the way I at least saw it.
00:06:19
Speaker
Okay.
EduChristine's Business Model and Expansion
00:06:20
Speaker
Interesting. And was the model online or was it like, NIT was not online? It was like offline first kind of a business. So what did you want to do?
00:06:30
Speaker
It was not that we wanted to do, but the model started, you know, somehow for us in, you know, in both the formats. Because I think from day one, it was clear, you know, what is that scalability is going to be one of the key metric. But at the end of the day, what do you scale, you know, online training, as we know today, wasn't really established because getting even a broadband connection wasn't, you know, easy, a stable connection. So we started with, you know, I think,
00:06:56
Speaker
large part was in the classroom and we were always running from day one we were always running something online you know and slowly the percentages you know kind of increased for online versus offline that's how typically you know we moved about it which became eventually a blended model but the thought that we are going to do it online was there from day one
00:07:21
Speaker
Absolutely day one. That's how we started. Yeah. We were the first ones to sort of start using online training. So we use a software called Dim Dim, which was an open source software. That's where we started. So most of these softwares are not even stable at that point of time. So, so I think eventually we found Citrix go to webinar or go to meeting to be the first stable software in this space. But when we started, there was practically no other software that was available that was
00:07:51
Speaker
sort of stable and affordable. When you say part of it online, does that mean that it's a hybrid course or does it mean that some admissions are for an online course and some admissions are for an offline course? Both. So we had purely online courses also. So that means people would enroll just for the online course, they could come.
00:08:14
Speaker
This is right from the first batch onwards. We had people joining just for the online courses. We had people joining for hybrid. So we had all the three models. So that means people who could join live online courses, people who wanted asynchronous online courses, that means they wanted recording and then they wanted classroom and then there was a hybrid too. So all those formats were there.
00:08:42
Speaker
The interesting thing was, from day one, the prices of online courses were at par with the classroom courses. So we were very clear that we are not going to offer something which is like different or something which is an abridged version. It is the same thing, except that the structural medium is online. That is what we maintained from day one.
00:09:04
Speaker
Amazing, amazing.
Profitability and Funding Success
00:09:05
Speaker
So this is 2008 when you lost your first course. When did you feel that you have found product market fit? And what does it feel like when you know that you have found product market fit?
00:09:17
Speaker
I don't know, it's difficult to decide that. So for us, as you can see, it's primarily a services game, right? So that means on day when you start getting revenue, so you cannot run a batch unless you have students in the batch. So you can create a recording, but if you're running it on sort of a classroom, students have to be there in the classroom. So for us, the deja vu of
00:09:43
Speaker
of a student being in the class was in day one. And interestingly, sort of when we were running the FRM classes, there was almost nobody in the market at that point of time. So that means, and it was a very technical, high-end course. So that means we had students joining in right from day one.
00:10:01
Speaker
But you can say there is another sort of point where you feel it's an inflection point where it's just not like a few people joining in. It's like a corporate running. So I think that took a few years. That did take some time. That's only when we started CFA and financial modeling that sort of we thought that it was a stable revenue generating company that could live on its own. Okay. Okay. Okay.
00:10:27
Speaker
And did you need to raise funds to reach that point of inflection? Because as you said, education means you don't need working capital assets because students tend to pay in advance. So did you need to raise funds? And if you did, then what did you need them for? And just to add to Parandit's part, we started with
00:10:54
Speaker
one course and one city, FRM in Bombay. Within two years, we were in three cities and we had three courses. And of course, they're also running online. I think at that point of time, somewhere around that time, we figured out fine. We can run multiple courses, multiple cities, multiple mediums.
00:11:13
Speaker
We know how to essentially fill a batch of 30 people, 35 people. And somewhere we said, let's the time to kind of double down on everything. But you know, as the nature of the business is that initially when you try to run a new city or a new course or you try to build a brand, essentially you're required to do a lot of investments in the beginning.
00:11:32
Speaker
to develop a course, to develop a center. Of course, we were trying to keep it as a slide, but still it required that money. And that is when we essentially raised funds from Dr. Mark Mobius, who was a well-known person, writer and angel investor.
00:11:49
Speaker
and Rajesh Siegel, who heads Graninity Ventures. Both of them joined us as Angel Ventures and I think within a few months from that, we raised funds from Excel partners. And again, this was the time when funding wasn't as easy to get or it wasn't available.
00:12:10
Speaker
So I think that was in 2010. I think that's when we... We ran our business profitably for a couple of years. Eight till ten, I think we run it very profitably. And that's when we thought it made sense to invest and double down on the progress that we had. That's when we raised money to expand.
00:12:36
Speaker
What was, I mean, did you find it challenging because first time founders trying to raise funds, what was some of the learnings you got while attempting to raise funds?
00:12:48
Speaker
Getting a fund to really sign off was challenging. But the entire process and given that we were more than two people, we were able to do it in that sense much easily. It wasn't really a trouble for us, for everything. Be it negotiation on the term sheet, be it financial modeling, be it projection, be it slides. I mean, we were doing it anyways. It was okay.
00:13:13
Speaker
That's the advantage of being in the education space that you understand how to present, how to make the narrative. And we were eye bankers actually selling companies, you know. So we knew how to do that. It wasn't that difficult. So here's
Global Expansion of EduChristine
00:13:29
Speaker
a question that might be a little hard, but why isn't Andrew Pristine counted among the tech unicorns that everyone talks about today?
00:13:38
Speaker
Because you had an early start. You had that first more advantage. You raised a decent round by 2010-11. The momentum was there. You were in a space where probably I was giving high ticket size. I think there are two reasons. I think one, unicorns in terms of valuation tend to be very sentiment driven.
00:14:07
Speaker
So that means if there's investor sentiment in that segment, you could very soon be raising very little money and then still achieve a unicorn. So unicorn status. So that means literally you could be raising a million dollars at a billion dollar valuation and still be called.
00:14:24
Speaker
And still be called a unicorn. So that is one part. So it is very investor sentiment driven. So that means when people are investing in that sector, they keep investing and sort of the valuations keep rising. When we were there, there was
00:14:40
Speaker
I think Tutavista raised money in 2006-07. That was a large amount of money. They did not do very well. It was more of homework online, online homework. Online homework systems, online studies. With the India-US arbitrage. India is in India, customers in the US. So I think they raised money in 2006-07.
00:15:07
Speaker
16 and that's not a lot of money from today's standard. But from that time started, it was a lot of money. I think there is 70 million or something around that. They did not do very well. And then 2007 till I think almost 2018, there was no activity in that space. There were obviously investments. People were doing well.
00:15:29
Speaker
From the beginning, we were very conservative founders. Even though we raised money from Axl, we never burnt a hole in their pockets. We were always very conservative in the way we expanded. It was mostly profitable expansion since the beginning.
00:15:50
Speaker
or even at the losses, they were like minimal losses. So I think founders who tend to be conservative, they are not able to raise boatloads of money to burn. They do not have that chance to burn money. You can differentiate the founders who have that DNA and those who don't have that DNA. So 2011, you had raised your seed round from Axel and some angels from there.
00:16:14
Speaker
Yeah, it was business as usual. So we would run a very profitable business. So that means all of our batches had to be profitable or if not profitable, they need to make sense strategically. At some point in time, we should have seen profitability. Then we expanded a lot. So we expanded a lot of courses. We expanded.
00:16:31
Speaker
It was a lot of geographies. So we did very well in Middle East, Africa, Southeast Asia, the U.S., the U.N. batches. And most of this growth was profitable. So we needed to invest some money, get that running. But once we had that running, it was profitable.
00:16:49
Speaker
So I think that's why we moved from, let's say, a pure VC fund to a private equity stage within five years. That's what most of the companies do. But we did not raise more money. So we just raised series A and then we directly sort of were able to get to a private equity level. And when we were sort of looking at the private equity level, that's when
00:17:16
Speaker
We got interest from a very good strategic investor, so we thought that it made sense. But we still had a private equity investor and a strategic investor investing in us in 2016. So we went to a larger, largest company in the education sector from that standard.
00:17:35
Speaker
These five years sound to be quite a ride. I mean, you've compressed it into a few lines, but I have a bunch of questions. How did you figure out going global? What was your strategy of go-to-market and what are some of the learnings you can share about going global?
Transition After Acquisition
00:17:52
Speaker
Why we went global because we were doing online classes since the beginning.
00:17:58
Speaker
And we had had students joining us from outside. And we soon figured out the courses that we were offering, maybe the CFA program or the financial modeling program, they were universal in nature. Unlike typically HR or marketing versus a very, I think, local. And these are international exams. The audience are joining us from all across the globe. They were, of course, we were charging less. And what we figured out was these guys are more than willing to go online, pay us more.
00:18:27
Speaker
And beyond a point, once you have exhausted the cities in India, and you know that fine, you have got the market over here, you can always grow over here without any problem. Then we decided to really incorporate a company in the US and really go all in. So it was like, let's go all in because otherwise you really can't get that growth. And we had examples of companies like Kaplan or Trading the Street, which had done this.
00:18:51
Speaker
Wall Street prep. And by the time we went to Google, I think we were the largest training provider for these international exams in the world when it comes to absolute numbers. And the idea was fine. We knew the formula.
00:19:05
Speaker
And the formulas also, you know, I think let us just fill out the formula for this, you know, what was the formula? The whole formula of education was getting the right teacher in the classroom. And for that, our model was we'll get the right professionals conduct these classes over the weekends.
00:19:23
Speaker
So by the time we went global, we had around 100 classes running every weekend by professionals from the industry. These are really top-notch guys in finance. Somebody who's actually a bond trader would come and teach you bonds. Somebody who's actually an equity portfolio manager would come and teach you valuation.
00:19:42
Speaker
That is something which we figured out how to do that in India. And even from abroad, even if we had to run a class in Kenya, we had to run a class in US, we had to run a class in Australia, we knew the formula, how to find the best guy for the job. Once we figured out that, yes, we know because the teacher was the main element, we could find the best teacher, the court material thinker standard. We were authorized training providers for all these exams. We thought,
00:20:07
Speaker
Fine. Why not do it outside? And that's how essentially we just went all in and went global. And you would recruit teachers from that geography? That's correct. That's correct. What was the turnover from 2011 to 2016? What was that movement like? Your ARR.
00:20:33
Speaker
Rough numbers. I think 2011, we were around half a million. We raised our fund, about half a million, and profitable. I think when we raised private equity, we were around 6 million. I think more than that. 2016, what happened? Then you got a strategic investor, like an acquisition?
00:21:02
Speaker
So we were actually sort of selling products. We were the largest training providers for CPA. So just like in India, you have CPA, in US you have CPA. So we were the largest training providers for CPA in India. And there's a company called Becker. So we were sort of working with them because they had the best content in the market.
00:21:23
Speaker
So we were selling their content, we were using their content for our training. And from the very beginning, we had always made a point that we would always use only licensed content, or we would create our own content and then train people. So we were sort of licensing their content and then sort of using that content for training. That's when they approached us saying that they wanted to invest in us and take some stake in the company. At that point of time, they did invest in the company along with the private equity fund.
00:21:52
Speaker
And we at that point of time, I think these take that they wanted, there was no point getting that money just in the company. So the founders also part exited company at that point. Okay.
00:22:07
Speaker
And you had a period for which you had to stick around and what happened after 2016? Yeah, we were supposed to stick around with them and help them grow the company. And there was this agreement with them. So there was one shareholders agreement plus employment agreement. I think that's very typical in all the companies we had.
00:22:30
Speaker
So we worked with them for two years. Yeah. So we bought, I think these trading investments we got them 2016. And then we worked with them and we helped them sort of transition to the next level.
00:22:44
Speaker
Obviously, we already had a very good second level leadership because we had to build a profitable business. So right from the beginning, I think one choice that we made as an organization was that we wanted to build the right processes. So the company was not dependent on Pawan Meer and the co-founder to be successful. It had to run at its own. So at one point of time, we were almost running.
00:23:06
Speaker
a thousand batches in parallel across the world. So obviously no single person or two people can do that. So any other learnings you'd like to share from this stint, the first Edupristine stint?
00:23:20
Speaker
Maybe about org building or about how you bring processes alive. Creating a process on paper is different from bringing it alive. So my learning is that whatever market size you build on paper, that tends to be very different from what you actually realize when you get into the market.
00:23:42
Speaker
So our realization, for example, let's say, especially going to colleges, then trying to train students, that looks like a very huge market. Very, very difficult to crack because students are already very busy in schools, colleges. It's very difficult to get rid of them.
00:23:57
Speaker
So although on paper it might look like a very, very big market, it's difficult to realize and then capture that market. You tried that also, like building courses for college students. I think that's a mirage that everybody has to try. Okay. By learning, you know, there are two levels, you know, one especially regarding org building right and
00:24:26
Speaker
I think we were building a kind of community over there because we were dependent on external faculty. What I've at least understood is the more you are transparent and communicate with everyone, all the stakeholders, you know, this is what it is, you know, whether they are these, of course, people always, you know, want a policy for compensation for everything, right? So the more you support that with data and meet transparent people, then understand things.
00:24:50
Speaker
Of course, we will find a couple of guys who won't really get that. But the majority understands this. We really put it up front. Honestly, this is what it is. This is what we are doing. This is what we are facing. So how we plan to make changes, please give your suggestions. I think just talking to people and listening to them. Honestly, I think that essentially goes a long way in building an org. I think that's what my general sense is. Especially for people from outside. So I think that goes a long way in building a community kind of structure.
00:25:20
Speaker
Okay, amazing. So what next? I'm assuming that by 2018, both of you would have been reasonably well-off after the exit. You would have got some amount of cash also in it. What did you want to do next? Build something meaningful.
Shift Focus to AI and Machine Learning
00:25:39
Speaker
And I think by 2018, of course, we had kind of exited. We had not to worry about our next day about meals or anything that was taken care of.
00:25:49
Speaker
We were trying to find something meaningful. Tell me about that journey.
00:25:57
Speaker
I wanted to do something in technology. So again, both of us thought that you would work together. And that's when Pavan was based in Bombay. I was based in Delhi at that point of time. So we thought we would do something in technology because that's where most of the action was happening. It seemed to impact a lot of lives. So we thought that we would do something in technology. Again, we were not very clear because we had been away from programming for a very long time almost.
00:26:26
Speaker
For 15 years, we were away from programming. And I think if you're not clear, then the default option is to start with services and then see where it takes you. So if somebody has a problem, try to resolve that problem so that you can find your problem to jump to. So I think that's where we converged. What did you converge on? What kind of services?
00:26:56
Speaker
So primary technology, again, sort of at that point of time, I think artificial intelligence was just picking up. It was still not there. So, but machine learning in terms of building regression models, et cetera, was already there. So.
00:27:13
Speaker
We had a passion for numbers, so we had been doing numbers for a long while. We also started our courses around business analytics. That's where we thought we would start something in technology, we would start something in services, solve problems around numbers. Primarily around machine learning.
00:27:34
Speaker
And then very soon we sort of thought that a relatively tougher problem would be NLP. So that's where we sort of started getting deeper into.
00:27:46
Speaker
What was the problem that came in front of you? So did someone come with a problem that you decided to solve? Yeah. I mean, we were, of course, looking at problems like we also, you know, looking at our friends and people in the US, you know, and we, again, this was more like an MBA kind of thing. Fine. Know that we want to do something meaningful. It appears technology is the best way to go about it, right? Something which you are pretty much comfortable with. And we understood that machine learning, this whole space is opening up.
00:28:16
Speaker
And in this new space, we had equal chances compared to any other existing company. I think these are the kind of check boxes that we had in our mind. The other thing which we had also been into global markets, we also realized that it would be useful to try to crack and look at only one market. So from day one, we started looking at US. These things kind of started converging.
00:28:41
Speaker
we were able to get a client also very, very starting point, something just, you know, couple of people just to get a foot into the market. And while discussions, while doing this, we figured out, especially when it came to machine learning, there was this whole space of e-commerce and reviews. There are also one more space of automobiles and so on, which was very heavily to data and text. And there were more proven business models of affiliate marketing in this space.
00:29:08
Speaker
So we knew that if we're able to build something, it's not a wide loose chase. And if one is able to do that at scale, it will change things to that extent. So I think that's where the whole idea started making sense. The more we dug into it, we were able to solve the uninitiated, what is affiliate marketing? Amazons or flip cards of the world. They essentially want all the guys to come to the website and buy something.
00:29:33
Speaker
and up, they are more than willing to pay commission anyone who is sending traffic.
00:29:40
Speaker
towards them. If you are a good blog writer, you know, if you are a good, you know, whatever influencer, and if you promote the product and people go there. So Amazon in that case will give you one to 2% of the revenue of that product. This is known as affiliate marketing. And since if you look at the size of the kind of, you know, merchandise these guys sell, the affiliate marketing commission, the whole cork with itself, you know, should be of the tune of about $10 billion.
00:30:09
Speaker
is what it disbursed to companies which sent traffic towards them. Of course, a lot of this is built up by Google Ads or something, you know, you can say that, but it's a kind of well-known established large market. And the way we saw things over there was that this market was really fractured, you know, and then people were really taking a product, really using it and then, you know, giving one video or review, right? And this was not at scale. It was a good review. It was helpful.
00:30:38
Speaker
But it couldn't really be done for the billions of products. And that is where machine learning and AI would play a dominant role in this. And that's why typically this whole problem statement came into being. And then it went all in to solve the problem.
00:30:53
Speaker
So how did you build that at scale, like a way to get people to your platform at a pre-purchase level? As Perman said, see, a lot of times what people would do is like, let's say there's a blogger or influencer, he would try out a product. Let's say he or she would try out
00:31:11
Speaker
an iPhone and then write a review about the iPhone, saying that this is how I tried it out. It's kind of worked perfectly, but it lagged and heated and was very expensive. But overall, it's a very good quality product. Now, people who feel that this review was useful, they were
00:31:30
Speaker
let's say, if it's an affiliate marketer, they will click on the lane, go and then buy a product. If they buy the product and this commission is not charged to the consumer, then Amazon pays for this, for a smaller model, let's say 1% of the overall sales to the affiliate marketing person. So that's something that the platform pays reason being that they anyways have to acquire the customer. It could be through Google ads or they need to pay something. This could be through an affiliate marketing with the pay.
00:31:57
Speaker
So usually this market is all driven by these small bloggers who would sort of review the product, buy the product, or get it for free from the seller and then use it. And we thought that there was a lot of information already in the market. So that means a lot of people had already reviewed written blogs about products. And it was still very difficult to sort of assimilate this information. So let's say there are like, I don't know, let's say hundreds of people
00:32:25
Speaker
reviewing an iPhone, whom should I believe, right? And do I get to aggregate this somewhere? The only way this aggregation works is if you read and then form a view of what's happening.
00:32:37
Speaker
And sort of people do this manually at this point of time. So what we thought that there was a chance of utilizing machine learning and AI to analyze all this information, understand all this information, and then come up with the product recommendation. So we thought that there was no space to write more content because content was already there, but to ingest this content, understand this content, and then come up with a view.
00:33:01
Speaker
So, so as to help people by not getting them to read these millions of reviews, but sort of, we do the hard work, we do the heavy lifting for them. And we means the machine does the heavy lifting for them. And then we come up with the recommendations. This is what people are saying. Amazing. What you're talking about is, you know, even if there is an expert log.
00:33:21
Speaker
we can give more weightage to the expert blog and still take those inputs into our scheme of things. So it is slightly different in that sense that we were saying fine, we are assimilators and we are smart simulators to that extent. We know how much weightage do we need to give to old reviews, to expert blogs, what is happening when we were doing this. We knew from day one that if we are able to build this technology and engine smartly,
00:33:49
Speaker
Well, this is not a tool to be meant only for product reviews. It goes much, much, much more beyond that. And that's where the entire idea was. To start with one part, the product reviews, it goes to reviews at different levels, to travel, to hotels, to any kind of experience anywhere.
Creating a Product Review Platform
00:34:09
Speaker
But the whole idea was, can you really do it for a product? And there, the whole market was kind of thwarted to that extent.
00:34:16
Speaker
It's a tough nut to crack. Why is it a tough nut to crack? What were the hard problems that you had to solve at the back end so that the consumer gets that kind of a somewhat like a rotten tomatoes kind of an experience with like a recommendation that he can trust.
00:34:35
Speaker
So I think the hardest problem is scale. So see, doing it for a single category for a few products is not difficult. So obviously you can look at reviews and then come up with the recommendation. So the hardest problem in this scale, so that means, so we analyze over 10,000 categories, over 1 million products and over 500 million reviews.
00:35:02
Speaker
Now, at this scale, getting the technology to work, come up with coherent results, that's the difficult problem to tackle. So for example, let's say just going from one to five categories in terms of finding out what people are speaking about, that's a tough problem. So for example, for a guy with a laptop, people will be speaking about the screen, about the keyboard, about the RAM, and so on. But let's say if I were to speak about snowblowers,
00:35:31
Speaker
may not have ever seen a snowblower, I hadn't seen a snowblower.
00:35:36
Speaker
10,000 categories. So just understanding those categories is very difficult. So forming a view about those categories and then ingesting those reviews at scale for so many products. And that's a very challenging aspect of the product that you build. And most of the generative AI, again, at this point of time, tends to be very toy product. So you would see people doing it for a few sample things that's, I think, relatively easier.
00:36:04
Speaker
But getting it to work on scale and getting it to work consistently on scale, I think that's the hard part. And I think that there's one key learning from generative AI that at this point of time, at least you need a human in the loop for the quality to make sense. Generative AI
00:36:27
Speaker
It's a word I heard only after chat GPT was launched. I thought that chat GPT was the first mainstream generative AI product. Were you using generative AI back then? Because you were doing this around 2018-19, right?
00:36:42
Speaker
Yeah. So see, but generative as a term means that it just means that sort of the next token is a predicted token, right? So it's just not, so you are predicting the next token and then you can predict the next token and then you can predict the next token. So, so as it is text and token here means like a, like an expert part piece of text, next word or next character, right? So as such, this technology is in terms of understanding the actual language, this, this,
00:37:12
Speaker
this has been in existence for a while. It's just that recently sort of the quality has improved significantly. So for example, till chat GPT or till GPT-3, the quality of prediction was relatively very poor. It's just sort of the tipping point where the quality has improved significantly.
00:37:31
Speaker
We have been using NLP since the last four and a half, five years. And when we started using NLP, one of the problems that happens within natural language is vectorization. So that means going from text to a number so that next prediction is now very numerical instead of text, which is difficult to predict.
00:37:53
Speaker
So when we started, it was like literally a word to that means a word could be vectorized. That means the word could be converted into numbers. And now you have an intro fake or Microsoft coming up with models that can literally like.
00:38:11
Speaker
Create vectors out of hundreds and thousands of words that means and traffic has a model that can convert like a hundred thousand words into a single vector into a single number and Microsoft again tends to have a model they claim to have a model which could literally like
00:38:29
Speaker
be much more than that also. That means they can understand a single book in one go. So that's phenomenal. The models have become larger, but they used to be there for the last four and a half, five years.
00:38:44
Speaker
What was your product doing? Was it generating a review and a recommendation? That was machine-generated. If I click on, let's say, snowblowers, then I would get a generative summary and then links to explore more. We would analyze reviews and then rank and rate products. The primary purpose of the product was to
00:39:09
Speaker
let's say, understand the reviews of all the snowblowers that are there, then understand what people are speaking about, is it the sort of the throwing speed of the snowblower, what exactly are people speaking about, then sort of coming up with those features of the aspects.
00:39:27
Speaker
then figuring out in the reviews where people have spoken about these aspects because see people can speak about display a screen as nits as this by this as pixels, right? So there are multiple ways people can speak about, about screen. So, so each aspect can be like spoken about in English, there are like innumerable ways to speak about a single word. So then to figure out where people have spoken about those, those aspects, then figuring out
00:39:57
Speaker
whether they spoke positively or negatively about it, then going from that positive-negative sentence to an aggregated level of then saying that overall they spoke positively about the throwing speed of the snowblowers, that full pipeline we built.
00:40:15
Speaker
And then we obviously generated text also, we would summarize whatever manufacturer description of products was because we wanted to sort of standardize that summarization, we wanted to standardize specifications also. And it was very difficult for the last few years for the
00:40:32
Speaker
AI to do anything systematically and consistently. It's very recently that sort of the models have become more powerful, the computer has become cheaper, so you can do that at scale now. How does this work? Like at the back end of it, like say NITS is referring to the screen or that this person has a positive opinion or a negative opinion. Just so you know, given analogy on this, you know, how machine decides.
00:40:59
Speaker
So machines, you can imagine, they come in different flavors. So they're like, you know, class one student, you know. Imagine the machine is a five year old kid, you know, who studies in class one.
00:41:11
Speaker
This is what we were having around a year ago. You can only expect so much from class one or second grade student. Now, suppose there was a way to build a machine which is class 10 student. Essentially, which has read more books, which has already gone through much more. And that's how you see all this talk about parameters nowadays. 20 billion, 1 trillion, whatever it is. There are students now who have read more books. And when they read more books,
00:41:39
Speaker
What is right? What is considered to be right or wrong? Because they read so many stuff, right? Those patterns are already ingrained in their algorithms. Once they find something, right? They try to match and say, fine, I had read in so many millions of books, right? So that's what the machine goes about. That's the one way to visualize how machines work.
00:41:59
Speaker
machine learning algorithm. When we say that they are trained, you know, you can imagine that they are trained, you know, on this subject, they are trained only using history books. So, you know, that's why they are not very specialized. So, imagine the parameters of the grade and the training and the subject which they are specializing. You know, I think that will be a very good idea. So, earlier at West, we were dealing with, you know,
00:42:19
Speaker
grade 1 or 2 student flights. You can only hope so much. You have to fine tune them or talk to them. And they also used to give you more funny answers like my small kids would do. So I think that's one way to figure it out or feel this.
00:42:33
Speaker
When people started training machines, training them on 10 books or 20 books or Wikipedia as one matter, there's something called emergent capabilities. So the whole idea that once you train them on so much knowledge, their algorithm and their weights, their capacity to ingest so much data, we're still trying to discover certain new capabilities which almost seem like reasoning.
00:42:59
Speaker
So, for example, when you do, when you give a question of mathematics to a machine learning algorithm that divide, you know, 20019 by this, you know, and you get an answer, well, that answer is not coming from calculation per se. It is coming from a prediction of that number, right? I mean, this is again, we'll get you technical, right? But I'm saying, but there are so many other things which are kind of more emerging, which appears, you know, to be something else. But in reality, it is nothing but prediction of the next token or the word, as Pradip was saying.
00:43:29
Speaker
But those things are kind of emerging. And then now you can take input from this. You can do some fine tuning. You can alter the weights. Now, this is where companies like us are working and trying to maybe take it to the next level on the existing knowledge base, existing machine algorithms.
00:43:51
Speaker
For your product, the review platform, you were using an LLM. At that time, you said LLMs were primitive. But this was like you built your own LLM or this was available off the shelf or what?
00:44:07
Speaker
So when we started, we started with this LLM called BERT. So it was not a LLM, it was more of a vectorization sort of library, but we started with BERT. It was like a large language model. And then we, the first one that we started, that's actually the new parlance LLM that's called GPT Neo.
00:44:28
Speaker
So GPD Neo is like an open source version of the GPT-3 that we see from OpenAI. So we started with that. And now I think most of the new open source large language models that you see on Hugging Phase, they are like some variant of the Python architecture, which is the GPD architecture. So we...
00:44:52
Speaker
I started with GPT but that was again very recent so I would say that GPT we would have used last two years. Prior to that there was no GPT newer so we like literally built some of these models from scratch. So yeah so we use the word vectorization we use the sentence vectorization s-bertoid and then sort of build a model from there.
00:45:15
Speaker
What's the full form of GPT? Very basic question I have got. Generatively trained transformer. So, generative is that it generates the next token.
00:45:25
Speaker
pre-trained is that the model is already trained. So you don't need to train the model. It's already trained on a certain corpus. That means it understands some vocabulary. It has some corpus and transform is the architecture of the neural network. So each AI that you see, every AI that you see essentially is like a brain, right? So it has neurons and those neurons are connected.
00:45:45
Speaker
But what people found was that if you try to connect each neuron with each other neuron, the compute is just too heavy. You will not be able to compute anything at all. That means the compute needed is just not there. Then sort of people came up with multiple kinds of architectures that were there. So initial architectures like RNN, CRN, the sedoccurring neural network, or convolutional neural network, each worked on a certain
00:46:10
Speaker
type of problem. So that means there were specialized brains that could solve that problem. And now you have this transformer architecture, which was initially just for natural language. Now it is being utilized in other fields as well. So it's just like a certain type of brain where sort of neurons are connected in certain fashion.
00:46:30
Speaker
Can you go a little deeper on this? What does it mean when you say connect neurons? What does that mean? You gave examples that each had a different use case. What were those use cases? How were they different? Or is this too technical for someone like me to understand?
00:46:51
Speaker
See the way these neural networks work. So neuron you can think of, just like brain. So neuron is like a cell that computes something and then sees if there is an error or not. So think of your brain as doing something and then seeing whether this was correct or not. And then updating itself.
00:47:12
Speaker
So you can think of your, of your brain as let's say, I see, I see a color, the color is white, right? And I say white, then somebody comes with a feedback that is not white, this is actually green. So I update my brain saying that there's a new color now, which is called green.
00:47:29
Speaker
This is close to what I knew was white. But I update my weight saying that next time I see this color, it is green not white. Now just think of your brain as having multiple cells that do this. These neurons essentially give like a zero-one decision for questions. Yeah, they give a zero-one.
00:47:49
Speaker
And then they update themselves. They update themselves saying that sort of the eventual outcome is zero one, but, but before that zero come zero and you obviously see the color, then you interpret this color. So there are like multiple levels of decisions happening, but the final decision is a zero one. Right.
00:48:06
Speaker
So and you need a lot of neurons to give a decision and then you like wait it like you're saying that 80% of neurons are saying this is green. So it means it is green. Is that like you need a lot of neurons to get an accurate decision? I mean, why couldn't just one neuron be even before that? I mean, why couldn't we just use one neuron for this? Why do we need many neurons? That was the way I will say this is, you know, for example,
00:48:35
Speaker
If you think about a lion, the moment a lion can invoke fear, fascination, a particular image. So when you hear the word lion or when you hear the word bird, there could be hundreds of different kinds of birds that can come to your mind. And that's one single word.
00:48:53
Speaker
So, you can imagine that when you hear the word bird, there are 500 different bulbs in your mind. When you hear the word based on the context, 10 bulbs might be on, 490 would not be on. And suppose when you come across an eagle or a vulture, some more bulbs of fear or loathing, something would kind of become on. Now, imagine this happening for all the words that you have seen, everything happening.
00:49:17
Speaker
Now, imagine the kind of the bulbs that will happen. So essentially, you see this bulb, and you know the pattern forms in your mind at 5. When I see a vulture, these bulbs are on. And next time, when you come across a very gentle sparrow, and she comes and pecks your eye, your whole bulb system will similarly go for a toss for next time onwards. We'll update your bulbs. You'll update which one will trigger it on. So the way I see it in terms of neurons is, everything has multiple dimensions.
00:49:44
Speaker
Every word is multiple dimension. And these two put in conjunction words like n cross n cross n, it goes to different level. So, when we say an algorithm, it takes care of all these words, all these bulbs, all these words. For some algorithms, I can train them only on 15,000 words. So, they only build so much. So, they only understand so much as well. If something new comes up, they have to do some kind of approximation from there.
00:50:10
Speaker
So that's the way one can envisage neurons. Now, then another question would be, how quickly can you calculate those neurons in all those dimensions? So for example, a single word can be, must we look into 200 different dimensions? Can you compute that within milliseconds to really come to conclusion? How much time does it take for it to recognize something? Fractions of millisecond. So if you can do those calculations very fast,
00:50:35
Speaker
Then you start appearing like a human brain. That's the beauty of all these calculations. Earlier, these calculations would take minutes, so it would take five minutes to tell you something. Once it starts happening in milliseconds, the output that you see appears to be coming very, very human-like and they can do all those calculations and with powerful machines. I think that's one way to visualize and which bulbs are on, which are off.
00:50:57
Speaker
This is all what weights and equations. At the end of the day, it's just equations and vectors and, you know, matrices behind this. But I think that is one analogy you can, you know, maybe kind of visualize in your mind. I hope that helps. And how have neural networks evolved? Like, what were the early networks like as compared to the GPT architecture? Looked like in terms of sizes. I mean, how has it evolved? Like, why is GPT better than the previous generation of networks?
00:51:26
Speaker
The whole idea, I think, is about the transform architecture, which really, you know, transform the way things are being computed at a very fastest speed, you know. I mean, the way you do these maths and calculations, right? I think that is one of the defining changes that really happened. So this happened somewhere around 2018 or so, you know, when a paper had been published, you know.
00:51:45
Speaker
And from there, there were some more nuances in terms of translation regarding how do you pay attention to words and how do you use the transformers. And then those GPUs which were used for video games in terms of images, they in conjunction, because they could do metrics processing at the same time, at a very broad level. And a new kind of metrics calculation and metrics
00:52:09
Speaker
GPUs, right? One can do things faster. And of course, things move, you've already become better, the algorithms come into picture, right? I think that really changed the way things were. Can you make it more layman friendly? Okay, what is matrix calculation? And, you know, you use a couple of terms, which I did not understand.
00:52:29
Speaker
try to explain sort of how the networks have evolved.
Evolution of Neural Networks
00:52:32
Speaker
See obviously the first level of this of understanding brain is as I told you there are multiple neurons right they are connected with each other. So this is the simplest way to think of this is that connect each neuron to each other neuron.
00:52:46
Speaker
That's the first way to think about it. When we want to interpret the fear of an eagle, it has to be connected to a bird. A bird is connected to a flying object, but a flying object could also be a sparrow. Do you connect that to a bird?
00:53:04
Speaker
to fear or not. So that means everything is intermingled. Obviously, life is like that. Everything is intermingled. We are speaking to each other. We see each other's base. Everything seems to be connected. So that's the simplest way of thinking about the architecture. So each neuron is connected to every other neuron. And that's where neurons started evolving.
00:53:26
Speaker
Is that okay? So the neural network started evolving. Every time you go through an experience, you update all the weights. As Pavan said, you might love a small sparrow, but the moment you go through a bad experience, you update the weight signet. Okay. 99% of the time, this is okay. But 1% of the time you need to fear a sparrow also. So you update all the weights.
00:53:52
Speaker
The problem with this approach is that obviously in real life, we know each other, but we may not have met each other. That means there are some experiences that are still being left outside of this neural network. And the moment we try to sort of put all experiences, for example, I may have an experience on generative AI on neural networks, but I don't have an experience on construction, right? I try to build in this construction experience into me also. So the network becomes very bloated.
00:54:19
Speaker
and not all the parts of the network are being used every time. So you can think of that refinement as defining how to use the network and how to update those weights. So for example, the initial refinement like RNN, that's called a recurrent neural network, was used wherever you can see things that are recurring in nature.
00:54:42
Speaker
So people thought that sort of language is recurring in nature. So each time we speak something, we tend to then repeat that and then repeat that. So they tried to sort of fit RNN within that. But that tended to work, not always it worked. Then people thought that sort of images had like a convolutional framework. So they thought that you don't need to connect each neuron to every other neuron, but maybe to some neurons and then form a specific structure to that.
00:55:12
Speaker
And as Pawan said, in 2017-18, somebody came up with this network and this seminal paper called, attention is all you need. Is that okay? So on its set was very simple. When you are reading, there has to be context to reading. So for example, let's say I spoke to Tukpawan about webinar with Akshay.
00:55:39
Speaker
And then, we discussed what to say in that webinar. Now, we have an attention to power and me, and not to Akshay. So, how do we put giving an attention to Akshay? But at the same point of time, there are three people, so that means there are multiple heads of attention.
00:56:02
Speaker
So there is Akshay, there is Paramdeep and there is Bhavan and then there is another world called webinar, which is again equally important. How do you find all those words and then put a certain weight to all those words? I think that's where
00:56:17
Speaker
This paper gave a good understanding and then sort of proposed an architecture of neurons that are connected in such a way that this problem can be solved with that neural network. So I think that's where it evolved to a transformer architecture. Okay, fascinating. And why are the GPUs best suited for neural networks instead of regular CPUs? I think all the games and graphics, they required heavy processing.
00:56:46
Speaker
And they could, you know, look at because that photo trying to generate that whole pixels and pixels on a screen, right? And what needs to be updated, what is changing, right? So, GPUs are essentially designed for this. Unlike, you know, something which goes into linear format, right? Because they're designed to look at, you know, n-by-n matrix structure for all the processing, you know, the whole energy of bulbs and multiple dimensions, right? Then it turned out they were essentially better, you know, for such kind of, they can do much faster in parallel, multiple levels.
00:57:15
Speaker
If you look at a game, right? So when does the game look real to you? So the game looks real when each pixel of the screen changes on its own, right? So for example, you want a person running on the screen and then sort of trying to hit other person. You want all the pixels to update the same part of time.
00:57:35
Speaker
But if you see each pixel, the calculation is very small in nature. So for example, let's say if I'm moving, most of the screen still says white, it's just a small part of black that is changing. So you want those small calculations, but for each pixel you want them to be done immediately. So each calculation is small, but there are a large number of calculations.
00:58:01
Speaker
A CPU on the other way is like a heavy lifter. So let's say if I speak about an eight core CPU, it can do eight calculations in parallel, but all those could be large calculations. They could be very difficult calculations. So GPU enables you to do small calculations in parallel, but those could be like thousands and thousands of calculations. Whereas a CPU allows you to do heavy calculations.
00:58:27
Speaker
If it's an 8-core CPU, just 8 calculations. Fascinating. Okay. Good. I have a good understanding. I went into the weeds quite a bit, but thank you for your patience. Let's come back to, so you had this review website. Did it start making money for you? What, like, you know, tell me the journey from there.
00:58:45
Speaker
So when we started this business, obviously it requires investment. We are ready for all the investments that are required. So we are ready to raise funds for that. But yeah, so sort of the platform had to be sort of profitable. It had to have users coming in. So yeah, that's what our thinking was. And that's what happens when we build the platform. We had users coming in and sort of affiliate marketing started paying for it.
00:59:15
Speaker
If not from day one, maybe in a few months. And that is still running? Is that your main business? That was a test case to work out the technology. That is one of the products that we have. It is running. It's still profitable. It does very well for us. And it's a very good use case for us to showcase our capability around generative AI.
00:59:36
Speaker
And how much does it do revenue wise? Annual revenue. These days it keeps changing a little bit depending on how Google is treating you as in what kind of traffic it is giving you. I would say it has served around 20 million customers today. Okay. Or like what was the peak revenue? Just broad understanding. I think the monthly revenue we were doing around $80,000 a month.
01:00:05
Speaker
Okay, got it. Which is a good cash cow to have to support. But there are costs involved. This is the ARR, the revenue side. And this is not from GMV. This is actually net revenue. We're talking about not GMV, which everybody talks about. This is actual money that the company gets. You're one percent commission. That's correct. That's correct. So there we are talking about an ARR of one million. If you talk about GMV,
01:00:34
Speaker
Of course, the number is 100 million. That's different. So what next in your journey? Once you had this platform running profitably, what did you decide to do next?
01:00:51
Speaker
So we thought that most of these skills are generic in nature and a lot of large enterprises and businesses could benefit from these generative skills.
Applying Generative AI to Business Challenges
01:01:01
Speaker
So that's where our focus currently is. We want to use our capabilities around generative AI to help solve tough business problems. So that means that the accounting firm is looking
01:01:14
Speaker
to automate their processes, to get better results for their customers. An oil company is looking to understand contracts better, where these contracts will hit them. Let's say a media company is looking to automate their sort of campaign, creative generation. That's where we feel that this technology could have far-reaching implications and we would really love to utilize this technology to help our customers.
01:01:45
Speaker
Give me an interesting case study, something you recently implemented. I can talk about one regarding pedigree chart and maybe you'll find it interesting and I'll tell you why it is interesting.
01:02:03
Speaker
So typically, you know, when people go to a geneticist, right, to treat you for genetic rare disease or genetic disease, right. And this is a product that we have already built, you know, we are working with doctors on this. So essentially, if you go to a general practitioner, they would, you know, take some time for your history. What happened? What happened? And so on. But when you go to a genetics,
01:02:27
Speaker
Guy, they will take your entire family to be fine. You have this. What is the age of your father? Did he have it? Or mother? And so on. Half-brothers, step-brothers. And just interpolate this to a Western culture, where it becomes more complex. Essentially, where you'll have more half-brothers, and step-brothers, and so on. Do so far. And you require some intelligence to do this on pen and paper, or maybe on broad or something.
01:02:55
Speaker
What we have built is, and that takes time, 15 minutes, 20 minutes of the doctor, and this needs to be done right. What we have built is, essentially this is nothing but a decision tree diagram. You ask something, yes, no, go here, ask something, yes, no. That's what typically you also process this and make it, correct? If you have to use it through a program. What we have done now is, we have trained a large language model, right? We used a large language model.
01:03:23
Speaker
as a chatbot which can talk to a patient, actually infer all these relationships, build databases, matrices around it, and then publish the right graph over and around it. Now this is very different from a decision because there is no decision tree already around.
01:03:39
Speaker
It is slightly unique problem the way it is being solved. And we are hoping that this will start saving 15 minutes, 20 minutes of all the doctors. This is the kind of very small use case. But the way it works in the background, it's very unique. I think it's a great issue. How is it different? There is no decision to involve it. Imagine something like you talked about chat GPT.
01:04:05
Speaker
And you're talking about everything, my father days, my brother days half, and then it builds your matrix with write numbers and everything over there. Well, that's fascinating. How did you build this? What's the way to go about doing this?
01:04:18
Speaker
I mean, we take models, right? We have to build algorithms around it, you know, train the model, tell the model what to do next, you know, start giving ways to this. If you find this, do this. There's a lot of things called prompt engineering, that's things called fine-tuning, right? Everything goes into picture. I mean, that can, that is how typically, you know, we start doing it because again, it's like,
01:04:37
Speaker
class 10 kid you know and you have to train the kid how to draw a diagram like this every step has to be told every step has to be refined if you find this if you have to build a matrix put one if you find father and son you put two and so on that's you start you know building on stuff so that's how you go about it
01:04:54
Speaker
Do you need to train it with a lot of such sample conversations, like you need training data, right? So do you need to have a lot of sets? Yes, to make it better, you have to use lots of data. Again, there are techniques right now, wherein you can use some data and do something, a lot of things are called Crump fine-tuning. You can do Nuxian fine-tuning. There are no new techniques. You can also
01:05:18
Speaker
train a model to build you the sample data of, you know, and you can, you know, again, use that sample data effectively to train your original model. So there are things called reward model, which can tell you whether the sample data was good or bad. There are so many things happening. And again, that's where it becomes complex. It's no more a chat GPT problem that you ask something and get something out of it. That is only a toy use case at best, at best.
01:05:42
Speaker
Amazing. So you say you'll have one chatbot where you give it a prompt to create a conversation between the patient and the geneticist about the family history. And then you create thousands of such conversations and then those conversations are fed to the
01:06:00
Speaker
again it will create garbage to even to get to this you know again you have to train that but so at the end of the day you will require actual intelligence to really understand what is stupid and what is happening I think that cannot be taken away right now it cannot cannot so just one use gift is interesting I think another one is on a curriculum builder you know so again that's a very specific use gift for chat GPT that our company had built essentially if you go to chat GPT and say you know give me a curriculum of
01:06:29
Speaker
I want to teach this workshop on how to conduct an interview. It will give you 5, 7, 10 points and something like this. And if you ask more questions, it will give you more detail. But it will end over there. It will say fine, look at this. So what we had to build was how do you actually build a curriculum of 10-20 pages.
01:06:48
Speaker
in a PDF file with all the exercises, you know, a lot of things worksheet. How do you do that? So that is one problem which we have solved effectively and published a plugin on JAD CPT as well. Just to understand this, you know, this is not available publicly to anyone or in general in India as such, you know. So because one of our clients wanted to develop this, you know, in collaboration with us, we worked on that. I think that is, so there is one use case, you know, again.
01:07:14
Speaker
where we had built something. But there are a couple of more that we are working on, especially on e-commerce. How do you really do a virtual try-on of clothes? How does that work? So there are a lot of things happening over there. Again, there are some more projects that we are working on right now. And again, just to understand that part, putting on t-shirts is very easy. You want to put t-shirts on different people, that's still OK. How do you put a lehenga? How do you put a saadi? How do you put a kagra? That takes it to a different level.
01:07:43
Speaker
Everything is being done at some level. You can change the photo of Batman white to black, but if you have to change the photo of Batman white to black, that doesn't happen.
01:07:53
Speaker
Then you have to use and change the body type of the person, right? You are seeing Deepika Padukone in a sari and you want to see how would you look like in that, with your height, with your body type. It's really Prandipa 6-2. He would look something else. I am 5-4, I would look something else, right? So the entire thing changes when you start looking at a personal level. So these are very tough problems, by the way.
01:08:16
Speaker
So this would involve like the user uploading his picture or scanning himself with the phone camera and then the model would create a virtual avatar. Something like that you know again and the worst part is whatever you it's like you give a picture of yourself right now.
01:08:34
Speaker
And you have a picture of Paramdeep right now. And finally, you see yourself in the same Paramdeep situation but with your body type but with the same clothes. And in two-dimensional, there is no three-dimensional photograph. So there are solutions which require you to take a photograph from this angle left, right, up, down and then they can do this job.
01:08:54
Speaker
How do you do it smartly with some minimal intervention? These are some interesting problems that we are working on. There are two or three more different problems. Anything else you'd like to add? So there are multiple plants as I told you. So for example, these are the commerce sector and again sort of the healthcare sector. And also working in the education sector. Those three we understood very well. Finance we understand very well. So we are working with the
01:09:18
Speaker
the client who's looking to generate financial reports from the data that they have. That's more of a generative case. Then we're working with an auto sector company.
01:09:30
Speaker
Let's say like an equity analyst kind of a company does like. Correct. Okay. Bicyll recommendations in reports. Power. Yeah. So the recommendation is more sort of machine learning driven. The report around that is more generative AI. So that's one thing very working.
01:09:50
Speaker
But from that we're also working with the design product and like once you've built it for one time do you product is it and then offer it off the shelf to all of the clients or these are purpose built for each plant and that plant was the IP for it and like you cannot reuse it.
01:10:07
Speaker
So both the models work. So sometimes what we've done is that we have built accelerators. That means let's say 50% of the product is built and then you customize it for the client's needs. So we own the IP for the first 50%. We can reuse as many times as we want. But the eventual product, including the UI sort of is owned by the clients. That's one model of working. Sometimes we have built the product and the accelerator is sort of we own the product. Sometimes the client gives us
01:10:36
Speaker
Just a product to build that means they own all the content. What are the products which you own? So for example, Best Views Review has a lot of use cases where we own the product right from scratch. So within that, there are a lot of use cases that we own. So for example, we own a use case to extract structured information from unstructured text.
01:11:00
Speaker
So for example, manufacturer has given product details, you want to extract specifications from there. So we own all that code. Now, if you go and sort of pitch this to a client who is looking to extract this for the auto sector away, then we own most of the code that we could just be customizing it for the auto sector, fine turning the model a little bit for the auto sector and then
01:11:25
Speaker
providing them with the same. Similarly, we have built models that can summarize product
01:11:33
Speaker
Let's say product description is written by manufacturers. Now the same summarization could be utilized in, let's say, a financial report where multiple reviews or multiple views of, let's say, different analysts need to be summarized and then presented in the report. So we own, you already understand that code, we own that code, so we could just use that off the shelf.
01:11:59
Speaker
So, I think the model is... Just pedigree tree, we own that. Okay, you own that. Okay, okay, okay. Fascinating. So, this pedigree tree is something you would, like, it would be like a SaaS offering, you would charge a monthly subscription to geneticists who want to... Again, you know, there are different, again, as Paramdeep mentioned, right? So, what happens in a lot of these cases, especially if you look at, you know, medical and finance, especially in these two areas, you know?
01:12:25
Speaker
None of these agencies, or hospitals, they would never allow their data to be seen by anybody outside the systems. So the idea which will eventually happen, work out here is something like SAS might not work with any one of them. By the way, a lot of these corporates have banned using OpenAI in any form, in any form, in any of their systems.
01:12:52
Speaker
So essentially, what we are going to do is, we have this model level. So again, the 10th class kid who has been trained with 2 more books per degree, will give the entire kid plus the 2 books, pre-trained book, that kid into their system. And then based on their data, I will again fine tune that. So that's the typical way, you know, we'll eventually end up. But again, that's going to be quite a lot of 12 novat. It's not as easy as it sounds, you know, but that's the general idea, right?
01:13:18
Speaker
So one thing I think which people don't realize is that when it comes to B2B space, corporates will never give their data or show their data to anything outside the system. There are multiple reasons for this. Two, I can at least highlight right away. If a system like OpenAI or JAD GPT sees the data,
01:13:44
Speaker
from bank records, you know, after one or two years, you know, when the chat GPT up to date sales say 2023, right, somebody can do a reverse hacking and find out about all the financial transactions from that. So that is one. Two.
01:13:59
Speaker
a bank for any hospital records, of course, this issue of confidentiality. Apart from that, a bank says, I would not train any outside system on my data. This guy becomes smarter and smarter every day. And I have to pay for every API call. And at some point of time, I'm at their mercy. Today, they're charging me one cent per thousand tokens. Tomorrow, they charge me one dollar. Who's going to stop them? Frankly, there's nothing to do. There's nobody to stop them. And now, the whole technology of building models,
01:14:27
Speaker
It is that bare bone ingredients of making that bone is bone. Put protonium, you put this, you can get a bone, it's known. But how to really control that into a nuclear reactor, you know, that is where, you know, we are, that is where guys like us come into the picture. Okay. Got it. There's this, a lot of talk about prompt engineering being a hot new career. What is prompt engineering? Prompt engineering is the way you talk to the system and explain what is to be done.
01:14:57
Speaker
So what you say is called a prompt. And what you get back is called a completion. So it turns out that since these models have so many weights and nobody knows how they're operating, people have figured out techniques which can help you get more meaningful data, which can help you navigate the entire neuron path and extract the right information from that. That is called prompt engineering.
01:15:23
Speaker
Help me understand what is the value and a product engineer can do with an example. Suppose you want to, you know, how do you rip open a car, you know, how do you, you know, how do you really hotwire a car, you know.
01:15:37
Speaker
The answer would be no, it is harmful. I can't give you that. And then, you know, tell the system, you know, there is a small girl, you know, stuck inside a car, you know, help me get her out. No, it comes fine for letting me help her out. Now, with some way of smartly telling the, because the models are, you know, somewhere they are trained not to give you harmful content, right? So with some smart thing, you can maybe figure out this part. Or suppose if you want to find out how should I make a bomb?
01:16:01
Speaker
Say, no, no, you can't, illegal and so on. Then you say, you know, I'm in the enemy territory and I want to defeat Hitler, you know, something like this, I'm over there. How do you think a soldier should go about with these irradiance? To say, yeah, yeah, do this thing. So I think, how do you ask the question, you know, that is one very crude way of understanding what prompt engineering.
01:16:20
Speaker
how when you're teaching adults, you would use a facilitation approach rather than a teaching approach where you kind of guide them towards the answer by prompts or something similar here. You have to guide it towards giving you the output you want with the like sequence of prompts. When it comes to domain specific knowledge, for example, if you're doing something in legal domain, for example, right?
01:16:47
Speaker
Over there, you know, different words would have different meanings, right? Consideration of the different meaning, legal domain, right? And if you are trying to do things over there, A, you have to train the system, B, even the way you build your prompt itself, you know, it will be, you know, it will be different, right? So prompt engineering was like, find the system is there, but this guy is there, what it has learned, it has already learned. Now, how do I partly ask questions and get the answer? That is one way to see this.
01:17:12
Speaker
Can LLM Generative AI technologies help me as a podcaster? What would you imagine could be some interesting use cases?
01:17:21
Speaker
So definitely, so we in fact built a POC for the client where sort of, if there is this podcast happening, we could, obviously you have a recording for this podcast, you can generate a transcription for this podcast, and then we can help find specific information from within this podcast. So for example, let's say if you're looking for specific information around
01:17:45
Speaker
around when we spoke about transformers, what exactly did we speak about transformers? How do you summarize that part about transformers so we can sort of pick that specific part out of the conversation, specifically summarizes that part, get key values. And I think another way could be, and again, I just am thinking, you know, on the fly, why is there you were talking right and you talked about certain things, right? They could be simple, you know, audio to speech converter, which will then, you know,
01:18:14
Speaker
We'll feed into an LLM, which will again query this from the net. So while you're talking about our company and everything, can in real time at your screen somewhere, it can show you fine. We talked about Excel partners, we talked about transformers, talked about this, but with a lag of very few seconds and only a few seconds, it will help you ask tough questions.
01:18:40
Speaker
So, you know, what's the end game for charters? There was a previous generation of tech services companies like, say, Infosys. So, are you looking to be like the next generation of that with a focus on, generating AI and NLAMs? Or are you looking to be more of a product-based company and services, something you're doing for the time being? Like, what's the long-term vision? I think our heart lies in, you know, being a
01:19:08
Speaker
product heavy company. That's where the heart lies. And at the end of the day, we are able to create something meaningful in the beginning. That's where our heart lies, meaningful in technology. And we can understand that building a good heavy-duty technology product is very, very tough. It's very, very challenging. And as much as we might want it, let's see where it leads to.
01:19:35
Speaker
Yeah, like essentially in the long term you would like to have like one hit product which like then completely changes your trajectory and all the decades dream of that.
01:19:46
Speaker
Yeah. Okay. Okay. Amazing. And what's your head size like? How big are you currently? We are approximately around 220 to 225 people. And I think that's a big team. If you ask the question after one month, I can tell you the number is 250, right? That's how you can say that. Yeah. That's the way we are hiring for 250.
01:20:09
Speaker
And that brings us to the end of this conversation. I want to ask you for a favor now. Did you like listening to the show? I'd love to hear your feedback about it. Do you have your own startup ideas? I'd love to hear them. Do you have questions for any of the guests that you heard about in the show? I'd love to get your questions and pass them on to the guests. Write to me at adatthepodium.in. That's adatthepodium.in.