AI Fundamentals For Starting A Marketplace Workshop

Welcome to our Everything Marketplaces workshop, which is going to be on AI fundamentals for marketplaces. So of course, seeing one of the biggest technology shifts with AI, so we're going to be covering how this impacts marketplaces and fundamentals for leveraging AI when starting a marketplace today. So Vijay will be leading the workshop, and most of you probably know him from his past group chat, but if not, he has some real incredible experience leading the AI and ML team at Thumbtack. and also a typical academic lecture on the topic. So as a quick note on the format before we jump into things, our workshops are once again 20 to 25 minute presentations that are followed by a group Q&A. So if you have any questions throughout the workshop, you can post in the Zoom chat, then also use the Zoom raise hand functionality once we get to the group Q&A after the presentation. And that way I can call on you to come on and ask a question directly. So with that said, thanks for joining us to lead the workshop today, Vijay, and I'll let you start with the brief intro and take it from here.

Hey, thanks, Mike. That was a wonderful introduction. So I am Vijay Raghavan. I lead the Applied Science, which is the AI and also the ML infrastructure teams at Thumbtack. Before Thumbtack, very similar roles at Roku, and much before that, very similar role at Chewy as well. In today's session, I'm going to do slightly— I looked at other presentations that people have done, other workshops that people have done in the past. I want to do things slightly different, which is I want to I want to talk a little bit about how do you transform a marketplace through AI today? Or in other words, to ask a different question, if I were building a marketplace today, how would I build it in an AI-native way? That's how I want to approach it. Again, it's a very broad treatment of AI. I'm happy to go into some of the details in the question and answer session. at the very end, and we can kind of continue the discussion later as well. So with that, I want to start with the agenda for today. I'm planning to do these, I'm planning to cover these 4 topics. Number one is just so that everybody is kind of at the same base level, I want to talk a little bit about the fundamentals of marketplaces in general. AI or not, these are the economic fundamentals. These have been there for hundreds of years. So I want to make sure that everybody's on the same place. And number 2 is I want to talk a little bit about the business model. And then number 3 and 4 are on the operations and then the strategy side. So on the fundamental, so anytime when we think about 2-sided marketplace, it has a couple of universal truths to them. You know, it doesn't matter if it's marketplace in the US, India, China, or it doesn't matter if it is in the marketplace of content or ridership and so on. And it can be in a different business or a different part of the world, but these, there are some fundamental axioms of truth for marketplaces. Number one is there are two sides of— there are two sides to a marketplace, right? So there is the supply side, which is sellers. In the case of Uber, that would be your typical drivers. And then there is the demand side, buyers. In the case of Airbnb, for example, that would be guests. One of the universal truths is the value of a marketplace increases Along with user participation, right? So the more drivers and then the more guests that you have, the more drivers and then the more, the more sellers and then the more buyers you have, the value of your marketplace increases. Most of the time, the revenue on the revenue side, there is, there is this universal truth that, you know, either you're making revenue from the buyer side or on the seller side or sometimes from both. So the revenue part is by facilitating the interactions of the transactions on either side. All that being said, the complexity of two-dimensional or two-sided marketplace is that there is a lot of coordination that is required on both sides. For example, if you are an Airbnb, you want to make sure that the property that you're kind of offering is according to some kind of specification. There is some base level that is expected that you can kind of fulfill through an override, so on and so forth. Safety is also a concern in many of these marketplaces. So there is a lot of coordination that is required on both sides. And I feel on all of these 4 areas, AI can kind of help. And then we will kind of look at it in a little bit more detailed way as we go. So in my mind, there is no universal way of categorizing the types of marketplaces. But in my mind, there are 3 ways of thinking about marketplaces in general.

Number one is, a transaction-based marketplace.

You can kind of consider Amazon, Uber, and Airbnb as transaction-based marketplaces, meaning a third-party seller comes into Amazon, he finds a buyer through the platform, sells X, and then Amazon takes a percentage out of that transaction as the revenue cut. Same thing with Uber, same thing with Airbnb. So these are kind of transaction-based marketplaces. And of late, There is also this concept of content-based marketplaces that could be your YouTube, your app stores, your social media, which is TikTok, Instagram, so on and so forth. Here, similar type of behavior where there is a buyer and a seller. Buyer, you know, buyer is the person who is watching the content. Seller is the person who is kind of creating that content. And content-based marketplace make revenue 2 ways. Number one is there is this payment kind of workflow where you pay X to Uber, I'm sorry, YouTube, and then that revenue is shared among the content creators, or there is an ad component to it. And then that is also a revenue share with different percentages depending on the platform. TikTok has a very similar revenue structure as well. YouTube is one of the first people to come up with this type of a revenue sharing agreement. And then you have professional marketplaces where basically it's a B2B platform for services, or it could be a B2B platform for goods as well. So I call it professional and B2B, professional when it comes to services, B2B when it comes to basically goods. So you have Upwork for services, you have Shopify, and then you have things like Alibaba, AliExpress, so on and so forth in terms of B2B. kind of marketplaces. Fundamentally, it is very important to figure out which type of marketplace you are in because the rules of the game are going to be completely different for these types of marketplaces. For example, if you are on a B2B type of marketplace, if you're on Alibaba, then the workflow for that, how you fundamentally build that using AI, is very, very different from, let's say, a professional marketplace or a content-based marketplace. As we go about this discussion, I would say one of the first areas of clarity is what type of marketplace that you are in, because not every rule or every suggestion will apply to all marketplaces. There is the generic thing, and then there is a specific thing that is based on the marketplaces that you're in. So one thing I kind of alluded to this before, but I want to make sure that people understand this. The biggest power of a marketplace comes with the network effect. The more suppliers, there is better selection for the buyers. Better selection means more buyers. More buyers means there is a larger supplier market. So this is almost like a— this is almost like a a wheel where, you know, once you have more suppliers, you'll kind of invite more buyers. Once you have more buyers, that'll kind of lead to more suppliers. So this type of workflow or this type of self-fulfilling prophecy is very, very important for a marketplace. And most of it is because of the power of the network, I think. Here is where I feel you can have a tremendous amount of value add through AI. What I mean by that? So, at the end of the day, one thing I want to kind of close the conversation and then we will talk about the role of AI here. So, all of this leads to a very competitive mode and then growth. Today, I don't think you can kind of question Uber in terms of ridership if you're kind of presenting plan. Let's say like an Uber to some VC, the question always people are going to ask is, how are you better than a Lyft and Uber? So understanding this type of network effect and then also the switching cost. Let's say you want to compete with an Uber or a Lyft, you need to worry or think about the switching cost. What do you offer that is much, much, much different from an Uber or Lyft in terms of a network effect for you to attract both customers and then Both riders and then drivers, so on and so forth. So one of the ways to think about network effect is identifying the right market, the right opportunity, and also a right way to acquire both customers and then suppliers as well, because The same tactic that you use for suppliers might not be the same tactic that you want to use for your customers, right? So both the things can be slightly different. Even if you're going to use the same marketing channel, let's say TikTok for one or YouTube for the other, the primary way that you kind of need to conceive these campaigns, need to think about these campaigns, need to think about how do you kind of spend your money on these campaigns to acquire those customers. are going to be fundamentally different. I feel this is where AI can play fundamentally a huge part. So I don't know how much of you have seen or thinking about the announcements from Meta even in the last 30 to 60 days. One of the things that Meta can do for you is not only they can kind of identify the right markets for you, they can identify even— they can even create the ads in the long term. So that's what Meta is talking about, which is they will help you to create ads, they will help you to reach right customers. They will keep, they will give you an idea of how your kind of campaigns are performing. So these companies, I would say Meta, Google, and then lots of other, what you call demand-side platforms on the advertising side, are going deeper into an AI-based ways or AI-based tooling to increase your kind of network effect. You can go to Roku and say, hey, this is the type of customer I'm looking for. And Roku has a DSP component as well. And then DSP is demand-side platform where publishers come with the demand inventory and then Roku will kind of match you to a particular channel, a particular type of ads, a particular type of context, right? If you're a Lexus, you want your ad to be shown after the race in a movie. So Roku can do all of that. Netflix also has an ad capacity today, and Amazon also has an ad business. In fact, if you look at how these businesses are growing, for example, for Amazon, the ads business is one of the most profitable and then growing business over a period of time. The reason I'm spending a little bit more time on the network effect is This is the fundamental way for you to create a long-term sustainable moat as you kind of think about platform. And, you know, we all think about word of the mouth, but word of the mouth, you know, word of the mouth is such a unicorn in this business. So you want to make sure that you are not dependent only on word of the mouth. You need to think strategically about how your advertising, slash how your content slash how your channel strategy could be much more effective in terms of building power that you can kind of create over a period of time in terms of network effect. And then how do you kind of use that as a competitive moat and then grow in the long term? So I would say use your partners to the fullest extent when you are talking to the YouTube people in terms of marketing and ad spend. There is there is an incentive for YouTube as well. So you are not powerless there. You know, you can be a small or a big player. Also understand that YouTube and Meta and the rest of Google, Roku, and all of these people have an incentive for you to use their platform in the long run, and then they are also rooting for your success. And then the question is how much you can kind of leverage their expertise here as well. So don't be afraid to ask questions to your YouTube. Rep or your Roku rep sales team or whoever you're kind of interacting with and ask for best practices, ask for more time to help you understand how the campaign works. So never be intimidated by these teams. Most of the time I've sat on the other side of the advertising business where I'm like, you know, you have so much power, to ask questions and then move that needle forward, why are you kind of not pushing it, right? So, the reason for all of these reasons, increasing the network effect is very, very important. And, you know, you can do it direct, indirect, you can get a ton of third-party data, then you can do a ton of social media buying as well. But, in this day and age, social media slash Performance marketing is such a big deal for all sides. It is, it is in Meta's best interest to make sure that you are also successful. So don't be worried to push that envelope. Don't be worried to ask more questions. Don't be worried to ask for more credits even, right? Most of the time it is, it is something like, you know, ask you shall kind of receive with these companies. So don't be afraid to push the envelope there. Okay. So, the value creation, we built a framework. So, in terms of value creation, if you think about a framework, right? So, you have the supply side, that is where, you know, you have access to customers, support services, and analytics and all that. And then the demand side, the value creation lies in how convenient is the platform. How easy is the price discovery? And then how much quality assurance you can kind of give. Most of the times in, in two-sided platform, on both sides you have this issue, right? From, from the demand side, as a customer, I come in, I need to think about convenience that you're adding, how easy it is to discover pricing for services, and what is your assurance for quality? So here also, AI can play a fantastic role, for example, on quality assurance. So even 18 months ago, making sense of video data was very, very hard. So now you have so many players in this area. All you have to do to assure quality is, let's say you have somebody do something X. And all you have to do is assure the customer that this particular work will be done according to this quality. And if there is any dispute, people can send you video, people can send you pictures, people can send you so many things. And then there is automated— the whole world of visual AI has changed a ton that you can easily assure quality. Same thing on the prototyping side as well, right? So even before you kind of go and create a product, there is so many ways to code and prototype these things that you can easily measure, or you can easily think about from a customer point of view, how convenient it is, how easy it is to kind of describe— sorry, discover price for a particular B2B transaction or a particular type of work and so on and so forth. So, there are so many things you can do on the AI side of things. Simply in these days, sky is the limit. The only limiting thing is your own imagination. Same thing on the supply side. One of the things that people have always told, or at least my experience has been, is when the other side is thinking about it, right? So when a pro is kind of joining your business or when a B2B transaction is kind of happening in your platform, the 2 things that people always ask for is data, which is the analytics, and any type of support services that you can kind of play. So support service could be a chat box to kind of answer questions, or it could be support service such as somebody, customer has a question to the seller and they need to reach the seller. They can do it by a chat box, a simple chat workflow. You can easily create it in about 10 minutes using AWS or GCP. Our Azure today that would have taken entire teams even 6 months ago. You can even classify chats based on tiers of severity. And then if your seller is in, let's say, outside of the US and then you want to wake them up in the middle of their night, you can connect it through PagerDuty. There is so many things you could do. So both on the supply side and then on the demand side, there is so many things that you could do on the AI side of things. If I were starting Two-Sided Marketplace today, I would kind of build all this natively in my app or on my website because this is what people, this is what people kind of expect these days, right? So people expect the ability to kind of chat their question and then get an answer immediately. In fact, I would say Human in the loop. Today, you know, people are used to self-checkout. People are used to kind of discovering their own answers. So, don't be holding to the assumption that there has to be a human at the end of the loop. Definitely, at some point in time, if a dispute is not resolved, you need to figure out at what point you need to involve a human. But the expectations have changed on the customer side as well. And on the pro side as well, where the expectation has gone from, you know, I expect a human to answer the phone on the other side to like, you know, I just want an intelligent chatbox that can kind of precisely answer my questions. So, you know, be aware of that change, I would say. The other thing I would kind of also talk about here is how you're going about creating this AI platform for this type of framework is also very important. Remember, today on AWS or on GCP or on Azure, you can create an end-to-end marketplace, I would say, in a matter of hours. And again, many of this, I told it in my last chat with many of you, but I would kind of mention this again. If you go to AWS, if you go to GCP folks, if you go to any AI vendor today, there is always a head of startups. There is always somebody whose entire job, that team's entire job is to make sure that you are successful. So, you know, they will build you prototypes. They will help you kind of come out with a minimum viable product as well. Because they also are invested in your success. They also will give you ideas how to make your chatbot better. Even things like search, out-of-the-box search. Back in the day, when I'm saying back in the day, it's not like 5 years ago, 6 years ago, it was 18 months ago. 18 months ago, to build a successful search box. It was a ton of work, but many of these players have made that very, very easy to do today. So, don't be afraid to talk to your AWS representative or reach out to the head of startup ecosystem. Most of the people have the title something like, you know, head of startup ecosystem or head of partner engineering and so on and so forth. Don't be afraid to kind of work with them. to kind of understand what a particular platform can kind of add value to. So a lot of these, for example, one of the important things on the convenience is finding the right person on the other side, on the supply side, who can kind of fulfill your request. Those are recommendation systems. And those recommendation systems have come a long way in recent times. If you go to AWS today or GCP today, There are services that you can kind of hook up so that you can build a very, very good recommendation system in about a few hours. Those were not available a long time ago. When I say a long time ago, I would say 12 months ago. So, the value creation framework is kind of, you know, having the right framework both on the supply and then the demand side. And then the platform is like, you know, you earn via ads, data, subscriptions, or fees. So, I feel, you know, a lot of times I feel founders/to-side marketplaces spend so much time on figuring out how to earn fees rather than— they spend so much time on the monetization side. And there's an opportunity cost of spending. I mean, monetization, please don't get me wrong, monetization is as important as other things. But the problem I see is that is the end result, right? So monetization happens because people on supply and demand side can find value in your platform and monetization happens as a result of that value creation. So a lot of times when I talk with founders and then when I talk with people in the startup ecosystem, immediately the conversation is on the monetization side. And then if you ask them questions like, okay, how are you thinking about your recommendation strategy? How are you thinking about your price discovery? How are you going to assure quality? I feel that is the area that if you kind of build the basics right, you can create a platform where you can have a mix of monetization strategies. And also people will have a lot of empathy if the monetization strategy changes even, right? So the first But, you know, the— for me, the first and the most important thing is what are you kind of adding value in terms of both supply and demand side? And what is your network effect? How big is your moat? Monetization is important, but it is not the first thing that you need to worry about. To me, I would worry about the supply side, demand side, and then kind of helping to build that moat. And then worry about monetization. And everybody is sympathetic to that thought process as well. If you build the right ecosystem, people will come on both sides. And monetization to me is important, but that should not be, how do I put it? Monetization should not be the first thing that you kind of think of. Because if you kind of think about monetization as the first way, then the problem is the opportunity cost of that is you will, you know, you will not be fully engaged on the value creation both on the supply and then demand side. So we have talked a lot about supply side and some of the things is, these are, you know, I have this slide for you just to think about the possibilities and this is not exhaustive by any means. So one of the things is when it doesn't matter if it's a B2B, if it's a transaction-based place or a B2B place or a professional-based to-sided marketplace. One of the most important things today is how are you going to expand the customer base of your suppliers? So one of the things that people are already talking about is personalized outreach. How can I create ads or how can I create content that will help me reach out to the customers in a more personalized way, right? So personalized ad creation, I'm calling it ads. It could be personalized content creation for different types of people, different segments. But how do you kind of do a personalized outreach to your kind of customers that thinks about the customer segments, right? So let's say you have people in Austin, people in Denver, people in Boston. And if you are a person offering services, let's say, you know, let's say college application services or something like that, the needs for all of these 3 marketplaces are completely different. Austin, much younger population. Boston, there is lots of institutions there. So actually the ratio of students in Boston is much higher than people who will actually apply for universities in Boston. And if you are trying to have a marketplace where I help students to get into the right universities or right colleges, these are much, much, much, much different marketplaces, how do you kind of make sure that the content that you kind of create for these marketplaces are much, much, much more personalized? So this is where I would highly encourage you to think about how Meta is thinking about it. One of the first things that happened this summer was there was a completely AI-created ad that ran in one of the one of our major athletic events this summer, and that was amazing actually. It was very, very personalized to the type of audience that would watch that type of a game. So personalized outreach, how do you kind of, not just in terms of ads, so it doesn't have to be a complicated video message. It could be as simple as a structured email message where, you know, I would kind of help people to kind of personalize that email message to reach the right people. Support services. How do I kind of automate the documentation? Service providers are mostly Spanish-speaking. How can I kind of reach them through my documentation? Do I have automated text conversion available to me? And I've said this before, but I'm saying this again, analytics is is one of the most basic things that your suppliers expect today. And there is ton and ton of easy ways to do it with AI today. You know, we have kind of talked about this before, so I'm going to kind of go a little bit faster here. So price comparison tool, price discovery on the convenience side, natural language search, personalized recommendation, instant answers. And then the most important thing is quality assurance. Don't ever forget that because at the end of the day, if the quality of your services through your marketplace is not up to par, people have very less sympathy or empathy for marketplaces that cannot offer a consistent quality. So you can have synthesized reviews through AI. You can have fraud detection through AI. That is where I was saying, like, you know, you can easily spot a fake review. Or you can easily spot through a video if a job was not done properly. And you can also have your own internal ways of generating quality scores as well. So there is so much work that has already been done for you. All you have to do is, again, imagination is the key here. So I said, don't think about monetization first, but monetization is one of the important things. You can do so much on dynamic fee structures. So based on the transaction patterns there are. So if you go to AWS, if you go to GCP, or if you even go to Databricks, there are notebooks for you where all you have to do is for figuring out your dynamic fee structure, all you have to do is take data from your transaction. Sometimes it is automatically hooked up that will tell you what type of a dynamic fee structure you need to have for different supplier and customer segments. A/B testing, automated A/B testing is very important. And many of these platforms also offer you this as a complementary service. In terms of advertising enhancement, ad copy generation, creative asset production is also something that generative AI can help you tremendously these days. So one of the most important things is how do you you know, overcome the bootstrapping and, sorry, cold start challenge when you don't have enough of a supplier or a buyer. So there also AI can kind of help you to kind of come up with different models on which type of geography/vertical focus you can do. How do you do community seeding subsidies? You can also do a lot of simulations to figure out many of these as well. So you can kind of think about synthetic network effect through AI. You can also, you know, you can also generate a lot of intelligent market data-driven kind of targeting as well. So this is not just not for you, but for your suppliers as well. So operations. So you have different kind of operating models. And this is where I want to spend a little bit less time because everybody knows this. You have a transaction-based model, you have a fixed-based model, you have subscriptions, advertising. So this is where analytics can help you a lot. But the fundamentals of what I've said in terms of value creation, if they are not there, there is no There is no amount of monetization play that can kind of help you. So AI can also help a lot in terms of understanding your unit economics. There is a separate talk by somebody from Poshmark where they went over a lot of unit economics. I would kind of like, you know, I'm not going to top that. So I'm going to highly recommend that video or that episode to you. I'm sure Mike has it somewhere. One of the most important things I've seen is people forget about CLV. Just in one single line, in most of the marketplaces, 80% of the value gets created by 20% of the participants. This is true for most of the marketplaces. And then there is a fair amount of academic research out there on that. And not all your customers and suppliers are created equal. So customer lifetime value is very, very important from day one to do it. So again, back in the day, 2 years ago, this is a ton of work, but there are companies who do this to you. All you have to do is give them your transaction data and they can kind of create a CLV framework for you for very, very less money. One of the biggest providers here I would kind of highly recommend is a company called Theta. The Theta, you know, I don't get any, you know, endorsement fees from them, but they are the people who created the CLV framework at Wharton. So it is run by 2 individuals who are highly respected in the CLV space. So don't be afraid to reach out to vendors in this area because CLV calculation is very, very hard if you want to do it. And especially if you're doing it for the first time, I would not recommend that. So they can do so many things for you. So, you know, in terms of predictive CLV, cohort analysis, and churn prediction, I would say that the AI-powered tooling has come a long way today. And you can use things like plug and play. All you have to do is have an arrangement with one of these vendors and they will use AI. Back in the day, about 18 months ago, There was nothing here. You have to create your own CLV models, and those are frustratingly hard to get it right because of the seasonality involved, because of all the other statistical complexities involved. But there are so many vendors in this space, and I would say check out data partners here. So, you know, one of the things is Usually the phases that people grow into a set of marketplaces, you start with something as free or with low fees, and then you would optimize it, and then you would add value-added services on top of it. The important thing is there are models that can tell you the price sensitivity of a particular customer, and also how do you automatically adjust search performance and then pricing and so on and so forth. Based on what a market can kind of take. And that is very, very different from each zip code even. So if you, you know, almost all providers, cloud providers have models that can help you with. The question is asking the right question and then reaching out to the right people to do that. So again, geographic expansion, category expansion, vertical integration, all is possible through a ton of AI modeling today. Again, I would kind of encourage you to ask the right question to your AI providers because a lot of this is already pre-plugged and set up for you. A lot of scenario analysis and planning tools, you can do it automatically today. So don't think about using or kind of coming up with your own ways of doing this because it is going to be incredibly hard. Especially when it comes to cross-sell kind of a deal. This is a very, very simple analysis that a simple scenario planner from any one of these cloud providers will kind of help you to bring that to your marketplace. Again, payments and logistics, even within payments, how to run your payments, right? For example, there are payment systems in the world. Payment systems are usually tiered, right? So the most expensive payment system is much, much, much more reliable. You need to run it in a particular cadence. So not every payment has to go through your most expensive payment systems because you don't have to have certain things for all your customers. So how do you kind of save money through payment systems are also a possibility. And there is lots of scenario planning that you could do around payments and logistics. Quality control. So one of the frameworks that I think about when it comes to doing quality control at scale is prescreening, monitoring, and then reacting to quality and then kind of improve. Again, community moderation and then facilitated by AI and kind of a tiered review could be very, very helpful. This is where agentic AI comes into play. So you can have agents look at your community moderation summaries and so on and so forth and take action automatically. It can block a particular vendor in your platform if the reviews are kind of scaling up and then they're not meeting some of the things. Like I said, all of this, AWS has a native agenting system called Strands that can kind of hook up your models and then kind of do this in a very quick way. So I would, again, nothing to invent. Most of these things are out of the box today. So I'm not going to put too much effort on governance. This is something that everybody knows. You do need to have an enforcement mechanism. Again, this is something that you can automatically do through AI. Again, In terms of competitive dynamics, in most of the markets today, it is a winner-take-all with a strong network effect. The thing about here is the switching costs. Your platforms should be so valuable that switching costs, both in theory and practice, should be your biggest moat, right? It's one of the biggest things that you need to consider in your moat. So, there are network-based modes that you can kind of create through AI automatically, operational modes, strategic modes. The playbook is moving fast and then having an exceedingly high velocity to kind of thinking about an idea and then moving to production. So, not so much of a comment here, but these are the success factors that you need to think, the traditional success factors that you need to think. applies to markets with or without AI. So these are like laws of physics, they're fine. Anyway, so in terms of— I will stop with one thing. The biggest thing that I want to think about here, I want to underline is the tech debt. So sometimes people don't understand the path that tech debt pays. You can move fast and break things, but the problem is understand that, you know, things are going to break. If you're going to put bandages as a stopgap solution over a period of time, things are bound to break at some point in time. So, one of the other side effects of moving fast is you will incur, you know, tech debt, and then you might end up having a bad trust system in your marketplace. So, think about the side effect. There is pros and cons for everything in life. So this is a con. However, if that is at an unreasonable or unsustainable pace, there's going to be issues. So that's all I was going to talk about. So I'm going to stop there.

Super comprehensive. So we appreciate, you know, walking us through not only the fundamentals, but also to, you know, specific ways you can leverage AI, you know, when I start starting a marketplace and scaling as well. So this will be helpful. So we are going to get into the group Q&A, so we'll stop the screen share right here. Is it, is it Iram?

And thank you very much, Vijay. That was a great presentation. So basically, I just had one question about something you mentioned around the supply side enhancement. So you mentioned that predictive analytics based on historical data could be an enhancement, and I can see a clear moat where that historical data is internal or proprietary. But when that historical data is something that's been sourced from the public domain, it's structured in a certain way. I was just wondering how much of a moat we can truly have given how good some of the LLMs are. Can someone use GPT-5 and get the same results as we would get with a well-structured internal dataset?

Look, the data that your marketplace is going to, going to generate is going to be internal to you. Right. So I would say that power in kind of setting up data collection mechanism that add value to you in the long term. ChatGPT can never have access to internal data unless you give it to ChatGPT. So I would say let's try to avoid that in the sense that, you know, don't— let's not create a marketplace where people are forced to download the data about their own transaction, put it into ChatGPT, then do the analysis, It is 3, exactly 3 lines of code to actually have your own version of ChatGPT API connection where you can send your internal data. And it is within, you know, it is within a setup where your proprietary data does not leak into the rest of the ChatGPT ecosystem. And then all the analysis you can do internally. So I'm saying is, Even on the analytics tooling, all you have to do is you need to enable the ChatGPT API and you need to have it connected to some kind of a chatbot system. And that can answer a ton of questions that people are asking about. You don't need to have anything beyond that, especially in your first set of iterations.

Thank you.

I'll actually jump in here and ask a question, another question that you, that you posted in the community, which is a great one and it's somewhat common here, which is around, you know, matching supply and demand. Since this is, of course, you know, core to kind of most marketplaces. But are there any specific scenarios or use cases that you see where AI— with AI, so how it can be leveraged to improve matchmaking and, you know, maybe some of those kind of like specific ways that it could be used?

No, definitely. So one of the first things I would think about this is when it comes to matchmaking, there are 2 things that are important. Number one is the capability of the seller and then the ability of the buyer. What I mean by that is In your matchmaking scenario, you need to be very, very clear about what are the things that a seller is kind of offering that is— that can kind of reduce the signal-to-noise ratio. The same thing on the buyer side as well. What are the things that can reduce the signal for the seller? Let's say, you know, I'm looking at— I'm looking at Mike providing some service. I need to understand a little bit about Mike's context, right? I need to understand a little bit about is Mike a price-conscious customer? Is Mike willing to pay for extra services that can kind of offer? So I would go one step further and say the more context that you can kind of create, it is better both for the buyer and seller. And in terms of matchmaking, there is so much work around this area natively around recommendation systems, I would say, that you can kind of natively use today. You don't have to build an entire recommendation system on your own. So first thing is, No problem can be solved if people don't understand what the problem is. Here, the signal-to-noise ratio of an insider marketplace is very, very high. And I would say, you know, that has to be clean and then that context has to be clear for both buyer and seller over the long term. If you have that, the technical portions again are much easier deal.

Yeah, we could probably go into more detail about that in specifics as well. So, hey, hey, Peter, I saw you raise your hand. Do you want to jump on here?

So if my predictive CLV model shows that a new client is likely to have a high lifetime value, what strategy should I use from day one to maximize retention? So make sure that your high-value CLV customers, right? So it is, I mean, one thing I will be careful about CLV is don't worry about the exact CLV. I would kind of classify your customers on deciles. The most important thing is make sure that you don't lose your customers in the top 2 or 3 So, that is your 20% or 30% high-value customers over a period of time. You don't want to lose them. So, in any strategy, just like medicine, don't do anything, you know, don't do anything to kind of screw up the current ecosystem or the current wellbeing of a patient. So, anything that you do, the most important thing is not to kind of drive those customers outside of your platform. So, I would kind of turn that in reverse and say, How do you kind of take the customers that you currently have, turn them into high-value customers, and how do you keep the high-value customers satisfied in your platform in a long time? So, you know, it takes 2 or 3 times more money and effort to acquire a new high-value customer. So a prudent strategy would be to keep up with the current high-value customers that they already have. And again, CLV is a little bit of a more complex, I talk whole day about CLV or even talk a week about CLV. So I would say when you are talking about CLV, it is very important to kind of get the right transaction data for predicting your CLV. You can predict CLV in different types of time horizons. So this is where I would feel like definitely involve a vendor or somebody who has kind of done CLV predictions for a long period of time.

That's a great question. Hey, I actually want to ask another question that a founder sent in advance, and it was around when building a resale marketplace. Do you have any maybe tips on, you know, after you kind of are doing a lot of that kind of like the manual kind of listing review moderation of like how you can leverage AI for that?

So the easiest way to leverage that is to have— I mean, it's a— there are 3 things you can do specifically. Number one is on moderation side, one of the issues is there is a ton of data and then people have a hard time figuring out what is noise and what as a signal. Like for example, let's say you have a vendor with 3 stars but they have 3,000 reviews and then you have a vendor with 5 stars but they have only 2 reviews. Obviously you need to consider the review that is done by more people, likely more, you need to create more weight for them. So you can do 3 strategies. Number one is simply you can have a summarized version of these moderated content. That an LLM can do. And basically it is a prompt engineering problem. You need to create a prompt and then say, hey, summarize it this way. And then you can even within the prompt, you can say, add this much weightage to the number of reviews, add this much weightage to the time period of the review, recent reviews, recent transactions carry more weight than older ones. So the entire thing can be arranged through prompt engineering. And all you have to do is connect this prompt to OpenAI API or to Cloud API or different things, and then you can kind of work it out. The funny part is you can easily ask ChatGPT to kind of even tell you a way to construct your prompt. So, you know, this is a simple prompt engineering problem. Again, don't be worried about experimenting it. So that'll be the first thing. The second thing is automating this process. And then the third thing is going one step further and even creating synthetic kind of idea to kind of get an idea of, let's say you have a new vendor or a new customer who came into your platform today. You can kind of, based on their behavior, even when they don't have any transactions, based on their screen time that they spend, based on other things that they kind of do within your kind of ecosystem, You can even go and predict what their scores would be 6 months from now, 3 months from now, so on and so forth. That is also possible through LLMs. So those are the 3 kind of things that I can easily think of.

Yeah, that's going to be a really helpful framework and kind of way to approach it. So yeah, it's been really cool. There's some cool ways as far as like how to do some of the predictive modeling that you mentioned as well. So cool. We're going to try to get in one last question if we can. I know we're almost out of time here. Hey, is it— hi, do you want to Oh yeah.

So I have one question. So actually, thank you for this one, all those AI applications. I think we need a big set of the AI toolkit. So I think one question is how do you put those together quickly? And then I think there are 3 categories, like which AI tools can be picked off the shelf from the cloud providers is the other question, right? And then which AI tools can buy from the third-party vendors, I guess, talk about CLV. calculation, and then which AI tools we should build in-house. So would that be a recommendation for those 3 categories?

Yeah, I would say something controversial. Hopefully it's not controversial, which is try not building too many things internally for 3 reasons. Number one is when, you know, when you're kind of starting The most important thing is you don't know what is working and what is not working. And then if you try to build something internally, it is going to take you 6 months. And then the chances of you getting better in terms of your recommendation algorithms that is better than AWS-provided recommendation or GCP-provided recommendation is very, very slim. So I would say when you're kind of starting new, the opportunity cost, there are things that you need to worry about. There are things you don't have to worry about. I would say the things that you have to worry about is how the customer experiences, how the, you know, how the supplier experiences. And then, you know, at the very end, the monetization, how your network effects are going to be. Those are the things that you worry about. And most of the success in marketplaces are going to be on those things. Fundamentally, AI, there is so— think about this, the amount of money that has been invested in AI tooling alone in the last one quarter in the US, it's something about like in the order of $350 billion or something like that. If you look at the CapEx spending of all our AI kind of providers, that is much more than the customer spending in the US for a quarter. So, you know, you and I, all 350 million people in the US, we spend less money than what the AI providers have kind of put on in terms of OpEx and CapEx expenditure in terms of AI tooling in the last 3 months or so. So, I would say, you know, for the most basic fundamental building blocks, if you're thinking about recommendation system, that is already there. If you're thinking about LLMs and then prompt engineering part, that is already there. CLD is one thing that I would say there is not that many good vendors in that area. So that is where I would say go for something like Theta Partners who have kind of done work in this area for a very, very long time. But other than that, honestly, you couldn't go wrong with any one of the vendors either on GCP or Azure. I would say in terms of, again, I don't have stocks in any of these players, so I'm going to say AWS is going to be the most responsive. GCP is going to be sometimes the most innovative. And Azure has got a leapfrog in terms of how they use OpenAI because they have such a big investment in OpenAI. So those are the 3 big things. But all that being said, in my mind, they're all very, very in equal footing these days.

Well, so that was a great question, a great way to wrap things up here. So we're almost out of time, but, you know, huge thanks for taking the time to join us here today and leading the workshop, Vijay. So you really packed it in and this is going to be incredibly helpful for founders. So I'll also be sure to put together a recap post to share the recording resources and also notes in the community. And so that way we can continue the conversation there in more detail. So thanks again for joining us and thanks everyone for the great questions today.