This AI Takes Down 5,000 Fakes Per Day ｜Jesse Xu @podqi

It takes years to build up a brand. It's very quick to loot them and they look indistinguishable now from the original be site. You can actually take these, put it into AI and that pop up these sites again again. It's pixel perfect. They're not just ripping off of this X hardware company. They're ripping off of anything from Bed Bath Beyond to supplements companies to random things. And so it's really like a network of counterfeiters, of infringers.

How big is this problem?

A two to 3 trillion dollar problem. It is very much a conflict of interest, right? By law, the platforms technically do not have a mandate to proactively take these down and we'll show them their results on the demo. Immediately, they see it and they're like, "Oh my god, I didn't even know there's this much out there. I had no idea. No human should be doing this work."

Hello everyone. Welcome back to TwoSetAI. I'm talking with Jesse Xu, CTO and co-founder of Podqi, who's taking on the trillion dollar counterfeit problem with AI. Jesse, thanks for joining. Can you tell us a little bit about yourself and your journey to build Podqi?

Thanks for having me on. I'll just give like a quick little background, but what we're doing here is we're essentially working up right up against the legal industry. And what we're seeing is that this is a problem that has been around for a long time, has been a huge pain point, but really it's something that I think every consumer is pretty radically familiar with now at this point. Essentially in a nutshell what we're working on is how do I take down fake everything, right and what do I mean by that is now if say you're a brand you might be getting from all angles so there was a whole the saga of like new culture being really big during the pandemic and only growing from there I think that's on the one end of okay here is what I mean by counterfeits in the traditional sense right where oh you have maybe fake product or unauthorized sellers that leak against the borders being sold being sold online. This has been around for a long time. It's a huge industry.

But really what we're seeing is that how this product reaches consumers is getting much much more complex. This is a combination of social media, ad buying. E-commerce is so much more direct consumer. I would say mobile is driving a huge part of that where people spend so much time now just shopping and seeing all of these ads, seeing all this. And really, it's okay. If I'm a bad actor, I there's a million ways that I can get to you, right? I can take out paid ads. I could put out fake domains.

I can send you phishing emails. And all of these are just crazy things to manage for brand. So that's what we're working on. We're essentially the protection layer against all these fake ways that the products can reach consumers. So we work with some top brands such as Netflix, Warner Brothers and entertainment industry. We work with some Fortune 10 hardware companies as well as some of the most popular consumer product brands. For example, something like a Jones Road Beauty, Fear of God, Mad Happy, things that you might be picking up to be recent Black Friday sales. That's that. For my story here. So I'm one of the co-founders. I started with two others. So Trevor and Ivan, also both my best friend. And really the progression was we were all in New York together pretty much doing our full-time jobs. And it came to a spark where the team came together first, right? We knew that ChatGPT had come out in 2023. This was just like the leading edge of that.

We knew we wanted to start something. We knew we wanted to build in a space. And really what kind of drove us is we wanted to look for the most boring manual and tedious space there was. And I had taken this class in patent law back when I was at Penn where I met Ivan, who was my freshman year roommate. And essentially what was part of that was the process of filing a patent of going to making sure that prior art and this is just one of the processes where you pay for essentially tens and tens of thousands of dollars to make sure that your patents are cleared.

There's no prior art out there and really it's just a bunch of paralegals in the back doing keyword searches and we were like this is crazy. It is something that's so painful and so tedious and that led us deeper into just exploring intellectual property as a whole. We pivoted away from patents and more towards the trademark and copyright side just because we saw that the volume there was much much higher. I always like to say trademark is that bedrock of so many crazy industries, right? It's like essentially the entertainment is all the soft IP, right? But at the same time, it's largest brands in the world. The reason why they're there is because they're protected by trademark law. And really, the reason why they can be those brands is they have to take the responsibility of enforcing on their brands. And so that's the problem that we really latched on to was, hey, how do you enforce? Currently, there's not really great solutions just pay lawyers that write letters. And so we want to dive deeper there.

You reminded me my days of working with patent lawyer and writing those patents. I feel like they speak a totally different language. Look at the different I cannot write this. Like this is my paper and this is the patent. It's unbelievable. It's hard to understand.

Yeah. And you're the expert. You're the one that came up with it. And you're like, I don't know what this is talking about.

Exactly. And I don't want to do it. I know. How long did it take you to pivot from patent to trademark in a counterfeit?

Yeah, so I think for us this was just around maybe 3 or 4 months. I think the real sniff test here was that for patents exactly what you said there is very esoteric. It has that sort of language where really it's almost like hard to get on purpose, right? Where they're purposely very much so a hey I want to have this way of speaking with this way of describing the patents that I think the sniff test for us was us three together we're not patent lawyers if we were for example to train an AI on how to recognize patent infringement for example we wouldn't be able to I wouldn't be able to tell you what is right or what is wrong whereas for trademark and copyright this is something where I would say the sense of seeing if something is trademark infringement is a lot better of a sniff test.

The real hard part is just finding all the ways that trademarks can be built upon. And essentially to backtrack a little bit, there's so many ways that trademarks can be infringed upon. Infringement is a very broad term. What we always say in our pitch is where the AI brand protection platform we find wherever your trademarks and copyright are being infringed upon. But infringement is broad, right? It can range from, oh, this is the one that's slapping your logo on TikTok. That's pretty clear. But it can range from even if they put or are associating them associating someone's logo on their website when they aren't actually an authorized partner or like also infringement, right? It's essentially disassociation. And so there's a range here where there's a lot of things that are very difficult and there's also fringe links where there are issues. Is problematic, but you're not going to be that one to one clear type please.

I'm going to ask you a new traditional question I'm trying to ask all my guests. You're the first one. Welcome. Yeah, I would love to start this question because I it really helps our audience and your audience as well understand what you're building in a very crisp way. So, can you explain what Podqi does at three levels? So how do how would you explain it to let's say elementary schooler a college student and maybe an industry professional let's say your customer IP lawyer

So breaking down the three level right if I were to tell what Podqi is to an elementary the way I go about it essentially what we're building is we help find how your favorite toys how your brands are being faked and ripped off. So, for example, one of our customers is Baby Shark, right? They're owned by Pinkfong. They're huge. You've probably heard Baby Shark. And the thing is they have their official merchandise. So, you gotten Baby Shark onesies, you've gotten Baby Shark toys, Baby Shark decorations. And the company makes a lot of that. And it's great stuff. It's very creative. It's very colorful. But there's also people that want to make fake version of that and they'll sell for cheaper and usually the quality is not nearly as good. And what happens is say your parents go out and they accidentally buy one of the fake versions of that. They'll be unhappy because it comes back maybe it's not nearly as good. It's not as colorful. It's not as well-made. And you might be sad about it. And then your parents will complain to Pinkfong and say, "Hey, I've gotten these goods.

They're not good quality. I'll leave a bad review. I'll call up customer service. When actuality, it wasn't actually Pinkfong's goods. It's not their fault. Except it is their fault that they made it, I guess, easy for you to buy those fakes. So, that's how we handle things. So, we find all the ways that are coming in online shopping things like that and we'll just take it all down so that you can make sure you always get those high quality things.

Wonderful college student

College student I think will bump up a little bit right which is okay to what I was saying earlier fake goods are goods are just one of the end of the spectrum I call that more of the clearcut end but really what we're looking at is infringement right and what we're seeing now is brands are being attacked so it's not just counterfeits and I think counterfeit has always been a huge problem. You look back, it's been a problem since the 2000s. Say Supreme was a huge one, right? That's a thing. You know, it's been a thing. It still is a huge problem, but now, you know, there's more angles of attack.

So really, for example, if I were a brand, there are things out there that are essentially a lot more malicious and a lot more priority for me now than maybe even counterfeits. So if I were Fear of God, for example, there's always Fear of God sweatshirts that are coming in from Express that are being sold and are bad quality and things like that. But it's how do they reach a consumer? How do they reach say someone that's shopped on Black Friday? Now you have essentially networks of bad actors that are putting up fake sites. So they'll fear of God.shop us. They'll take ads out on these sites and so they'll be on Meta ad library. They'll be on Google sponsored ads. And the nice thing about these is that they essentially are the top result, right? Those get served directed to you and they look indistinguishable now from the original be site. You can essentially take these, put it into your AI, and then pop up these sites again and again. It's pixel perfect. And so what you have is now you're essentially misleading consumers on these sites.

You're phishing them. In the most innocuous case, they might just be selling products and it never comes. Or they're taking your credit card information, they're taking that information and they're selling that as well. And so that's been a huge problem with brands, fake ads, fake sites, fake everything. So it's just getting more complicated now. And so for us, what we do is we have coverage of the marketplaces. We are a brand protection platform that really our bread and butter is those commercial usages. At the same time, we take down fake ads. We go to domains and take down websites and we essentially clear the board and make sure that always these things might be coming to consumers. We can essentially stop them in their tracks.

Professional

For professionals that next step which is okay for the largest enterprise brands so say Fortune 10 hardware this is all about the nuance right which is hey even I know that there is counterfeit out there of x good right but that's maybe not even the most priority thing right if someone is selling say a $100 good of mine for $10. Most consumers will know that's fake, right? That's okay. But really what the issue is, okay, there might be goods coming out there that are really indistinguishable from the real goods themselves. And so what is the next step to actually being able to really uproot these at the source? So this is where we get into investigative flows as well.

This is where a huge part of essentially our AI evaluations or AI search come into play which is okay I can pretty easily tell you all right this is say fake X coming in but how do I track where this is coming from how do I put it all together to say okay this might be coming from one manufacturer based out of some Asian country or this is the same entity that is putting out Facebook ads on X that run these profiles that's also linked in these websites. For example, you know, Loco and this other model that we see and we'll see that they're not just ripping off of this X hardware company. They're ripping off of anything from Bed Bath Beyond to supplements companies to random things.

And so it's really like a network of counterfeiters, of infringers. And so what we do is not only do we automate takedowns for our counterfeits for listings for ads and websites, we also trace back, grab all that information, add data and put it together so that we can help you with these investigative flows. And all of this is documented. So when we crawl through all this, we take screenshots as we go. We grab all information available and then you can use this for your evidence gathering. They say if you wanted to take retroactive compensation or go after settlement, this is what we can create that package for you and make it easy for you to essentially go after them and escalate things.

Is this something that you also work with the FBI and for reporting purposes?

Yeah. So this is like for example, we've chatted before with the FBI, but also with essentially like customs and border enforcement. So a lot of these goods, they do a great job stopping them at the ports. It's not going to be 100%. But really, that's a huge part of, hey, not only do we want to make sure that we can take them off online, but it's that boots on the ground enforcement as well. Hey, I know that X goods are very popular coming from this area. Making sure that those get stopped at tracks at the ports. They don't actually come into the US.

That is like an extension of the whole way of detecting fraud. I used to work on the fraud detection front for credit risk. So, sounds a bit similar.

Yeah. Absolutely. It's very interesting because there's so many data points that you can go after and really it's like almost sifting through a body and needle is that okay once I make a connection how do I ensure that connection is real and once it is okay boom it could crack open an entire case once I connect to

Who do you think should know about this the solution

I think there's I always bucket our best customer to two bucket on the one end are those enterprise brands these might be large Fortune 500 companies. They have existing brand protection teams or lawyers are dedicated to this. They've seen what's going on. They've been to a lot of the conferences like INTA that we go to and they really know, okay, this is how it's done. This is really the pain point. This is a problem. The real question is for them, okay, how do we improve our takeout efficiency? How do we improve on our investigative flows? Things like that.

So we bring in a lot more essentially workflows that automate the things that they do. So instead of having to gather evidence by hand, we help you do that. We crawl the web, find more hits and then with all those we annotate that data and then essentially help your processes that way. On the other end and I think this is something where we have to do a lot more essentially communication, a lot more awareness building is D2C is everywhere now, right? Like some of the biggest companies that we've seen recently were built directly on top of Shopify.

So this would be like say something like a groom or bridge wallet and these companies are the same way. As soon as we've seen pretty much overnight companies will crack say the top 25 on Amazon or they'll blow out on TikTok and then immediately afterwards fakes come in, right. So this be on say like TikTok Shop, fakes on Amazon will start rolling through. People will set up Shopify sites. It is so quick. It's like as soon as you become a target, they can just spin these up like clockwork. And the problem is for a lot of these brands, they're a very small operationally focused team.

They might be making tens of millions of revenue, but the margins are very close to, right? It's all about, okay, I'm here to make my product. I'm here to sell it. This is what I'm good at. This is what I want to do. Brand protection might not necessarily be part of the lexicon yet. What I do know though is I'm getting copied left and right and I'm seeing customers getting mad at me because they bought a fake when they got phished. How do I take care of this? What we did are is we're that plug-and-play solution, right? Because we're automated. We're a software platform that you can essentially onboard and in the first week start taking down fakes of your site of your brand. So, I think that's a huge part that it's really customer base that we're tapping into now and really growing for us.

So, I'm hearing you have the enterprise side of customers and then you also have the Shopify store owners.

Exactly.

The smaller businesses, but they're trending up. Maybe they're becoming more popular and then they become the target.

Absolutely. And I think that's the thing is just essentially for a lot of these bad actors, there's no differentiation. They just go after whoever is making a buck.

How big is this problem?

If you take industry reports, counterfeiting is growing year-over-year. As of 2021, it was estimated to be a $2 to 3 trillion dollar problem in terms of what are the most popular counterfeit goods. These will be supplements, these will be luxury, all types of fashion. But I think the problem is hard to measure in just the lens of counterfeits of these physical goods for example. I think really it's say maybe you're a top clothing brand. Let's call it Lululemon, right? Lululemon was dealing with such a big issue in previous months because of the whole thing of Lululemon dupes coming out of China, right? Of these factories that are putting out and saying, "Hey, we are making Lululemon in the background and we can sell it to you for a seventh of the cost." What we're seeing there is that is a huge issue in terms of your brand is losing value, right? It's not just the counter for goods where people will pick up Lulus for deliberately. It's more so okay now your reputation is getting tarnished. The awareness that your consumers have of you know how quality they believe your items are.

This is being attacked. And so I think extrapolating that this is more of what we want to essentially be able to safeguard which is it's not just the counterfeits and a pure price tag on those. It's more so the intangible hey what do people perceive of your brand right? And this comes from making sure that the brand is pristine, that they don't easily find counterfeits, that if I go on to Google in the first two sites, I'm not going to see fake websites. All these essentially is that cleanliness. And that's really where we're seeing people pay top dollar now is to being able to sh say, "Hey, it's right. I've gotten everything down." And we are known to be those guys that take brand awareness, brand protection very seriously.

That's also the intangible asset that's worth more than just losing the dollar amounts in the sales, right?

Exactly. Exactly. It's hard to put a price tag on it, but at the same time, you can feel it. And I think this is really where you'll It takes years to build up a brand. It's very quick to lose it.

That's such a great quote. And your positioning is very interesting, too. You talk about turning the brand protection from a cost center as well into a revenue engine, a revenue driver. This is a little bit counterintuitive. Can you explain more?

Really there's two end on one end is just as a baseline. What we've seen is that when we deploy with companies, we over a course of 6 to 12 months, you can expect to see around a 1 to 2% baseline revenue growth once we're in action and once we're taking down listings. I'll put it this way. Say you're at Alo Yoga Blue. I've really got this in my mind, so I'm just going to run with it. You're spending millions of dollars in ad spend essentially competing against each other, right?

Going after the same customers, making sure that hey, like we are the brand that fights it out. We're putting all these ads in front of you. Now imagine taking that budget and instead of competing with another brand, essentially competing against people that are selling fake goods, competing against people that are diverting revenue against yourself. So this will be essentially all those countless infringers out there that are putting out these fake products or putting out products that are so hard to distinguish that it's actually misleading consumers and really taking out a huge chunk of your pie. And so this is what we've seen is if we can clear the board essentially wipe these marketplaces off that be able to stop being purchased revenue grow.

On the other end, it more of the proactive fight, which is what we've seen for the largest brands and the law firms that we work with is they're not content with just sitting back and proactively clearing the board. They actually want to take the action directly to the counterfeit. And so what this means is that as we go through as we grab thousands of potential infringements maybe say 95% we just automate take down clear those from a Walmart from an Amazon but those 5% where maybe it's a particularly large entity a large amount of stock that's being sold or maybe it's someone that's particularly risky on a platform that's very high priority.

Those 5% what we've done is we go in grab screenshots of everything we document everything they export that our platform they hand it to the legal team. They'll actually go for retroactive compensation so they'll put a Schedule A where they will freeze accounts and essentially be able to grab all the any earnings that these sellers have made off their products or they'll actually get into these large suits against potentially even the platforms themselves because they argue hey you've not been doing a good enough job about taking fake products of mine off of your platform even when I've given you notice and these are quite lucrative, right? We've seen A lot of seven $8 million eight figure deals come out of this.

Yeah, that's crazy. Let's talk about the technical side of things. You mentioned about handling maybe more than 5,000 infringements per day for some of your customers. Do you have a technical architecture that you may be able to walk me through? Are you just scraping at massive scale or there's more to it?

Yeah, in a nutshell, the answer is yes. I always like to joke we're a scraping company. So that's where we put a huge amount of into essentially building our scrapers and building our architecture to account for that. What we've built is we're a distributed architecture. So I like to break this down into three steps essentially. It is okay, I scrape. So this will be all the largest marketplaces. This will be across the web. This will be image sources. This will be also like places, registries, secretary of state. So that we can find naming rights or contact information. We take that information. We then evaluate it. That's that middle stack. So this goes through and passes through. Okay, we have a number of classifiers. We match against any assets that we detect. We try to link to any existing entities that we've seen. And then finally it's that action portion of things. So once we've scraped once we've annotated and evaluated these results become our dashboard and then what we have is we always say okay we need a human in the loop sit there and do the action.

We always have someone on the brand's team or on a representative of theirs be on the platform and they are the ones that you know they know the brand the best. They are aware of okay who might be authorized who might not be and they are the ones to say okay these are maybe we want to ignore but say that 300 or 400 people listing you want to take action on we make it easy for them to essentially with a click of a button check those off send those all take down right and then on our end those are automations again so once you set it out then we take care of everything which is routing to the platforms being able to craft craft the justifications, making sure those go down, checking that they go down, things like that. So those are kind of like the three parts I really put our

Person who's approving all those takedowns are the person the teams with the platform. So like Amazon, not really the brand owner.

Really, it's the our users. The users of our platform is either someone on the brand side or potentially say for example a lawyer that represents them. Yeah. And then when we send these out for takedowns, ultimately it goes to say apps brand protection team and they decide okay is this essentially is this qualified is this justified takedown or not and then we essentially work with them. We craft things with them so that get those takedown rates possible but in the end our users are on the brand.

Then the user on the brand side are the ones who are initiating those takedowns. Do they go through any approval from the platform itself like TikTok or because this is against their interests, right? The platform will make money off these counterfeit as well. So, they're probably not policing it themselves.

Great question. And this is really the crux of why I think this is becomes a huge problem, right? And not to point fingers, but what we've seen is it is very much a conflict of interest, right? Because these platforms are making that essentially they're making margin off these products regardless of if they're on right and what we've seen is that two things one is some platforms are better than others at proactive policing right it is a thing where people do want to be known as hey we are a legitimate platform so there's always that some are more proactive than others and also in terms of actually enforcing this it's a whole different it's a whole spectrum essentially how efficient is it but I think the problem comes down to by law. Essentially, it is up to the IP owner, the rights owner to notify these platforms of infringement.

The platforms technically do not have a mandate to proactively take these out, right? Instead, it comes down to if a trademark owner or a copyright owner notified these platforms of the infringement, then all of a sudden the clock is ticking and these platforms have to act. But before then, there is actually no requirement for these platforms to proactively police. And so what we see here is okay in between these two interests there's a little bit of room where you know up until the rights owner reports the platforms are making margin off of this and really the hard part about this then is okay if I'm the rights owner maybe I see some things I fill out a form Amazon has their brand registry Walmart has their IP takedown form, TikTok Shop does as well, and that works for one-offs here and there. But if you have 500 infringements every day and it's just flooding the platform, it's never going to work. And this is where we come in.

You are integrated with these marketplace as well so that the takedown can actually happen without going through this lengthy process, right?

Yeah, absolutely. What I would say is we are a force in the end like you can think of what we do as pretty much what you could also do to the problems yourself. So you can manually fill things out. We have the luxury of a lot of times we have preferential treatment. We have programmatic access but in the end it's still very much that we report these for you. I think where the value comes in for us is we can do this at scale for hundreds and hundreds thousands of listings at once. And the second part is a big problem is also just finding these in the first place. You don't want to be looking up Amazon and Walmart and eBay and Temu and Shein and TikTok Shop and all these platforms at every one of their v variations all the time. But we can do that and that's what we'll do is we'll set that watch up. We'll run these reoccurring scrapes and then we'll find these listings so that you can actually

You have the data collection scraping piece. You have the machine learning pipelines that does the analysis and the matching and then you take the actions. That make total sense. I'm curious about your detection approach. Is there is it a traditional regression model or are you using unstructured data in a natural language processing? Are you using both?

So both the answer here. So we have a combination. There's your classes or classify classification tab right where it's very easy. Hey for example something like I need to tell if this listing is commercial usage or not. Right? If something is like us social media posts, it might not be commercial, we probably don't handle that. That's one example is that where we classify and then depending on that we might filter results, score them low and things like that. What we've seen be a huge boom for us is essentially the magic of that unstructured input into LLMs where you know while we're going through we are looking at listings from all the largest marketplaces but at the same time we're also looking at tons of new platforms of new results all the time. There's no structure to a lot of it right no site looks the same. And so what we'll do is before I imagine you pretty much had to either kind of hill climb and be able to say all right we have either models trained for specific platforms or we have some sort of generalized parsing from these.

And now what we find is when we get a new result a listener haven't seen a lot of our classification or a lot of our processes are able to just get this and evaluate this nearly as accurately as we have seen before and that's because we have the ability of essentially parsing in that input and that really being able to evaluate it without that structure. That's been huge for us.

So you'll be able to tell if based on the website information and the listing information is this a counterfeit or not or more likely because it's a probability right that's what you're reporting right

Yeah that's really the interesting part of this is always it's a stint right it's okay how can we find tells that say this is likely to be counterfeit price is the obvious one right if it's there's a huge price disparity that's very interesting this is very likely but that's pretty obvious this I think really where the beauty of this comes in is okay on one end right we want to make sure that your IP is being used in some way so you know we'll also have we have best-in-class image matching we have logo parsing but also it's that fuzzy matching as well so say if someone is taking your logo and obscuring some letters from it or they are distorting the image in some way we can grab all those and then that actually becomes a good tell for us hey if they're using that image and even better if this is an image that we've seen across a lot of our counterfeits for this client. Boom.

That scores is going on. But at the same time, there's always a lot of nuance as well that we get baked in. So, one of the things that we're most proud of is essentially we create what as part of the onboarding process, especially with these larger enterprise brands, is we will actually create custom profiles based upon the brand. So, we'll ingest their brand guidelines. We'll work with the team as part of the onboarding and we'll say hey you know what are essentially signs or signals that you've seen through your time working brand protection with this brand and your intimacy with this brand that would be great things for us to detect. And so this might be for example if this is some description or some snippet that infringers use a lot or they'll describe the product in some way or what have you. We can actually bake this in as part of our process and we essentially be will be able to curate this towards the brand's enforcement priorities themselves.

So that's been also a large part is okay then once we have all these how do I stack rank so that you have the highest priority items at the very top and then we scored that one.

That's really interesting. Are you training your own models towards each brand in any ways like the image models or the NLP models?

Yep. Exactly. So what we've is we have generalized models most start we can conduct these searches and we still can get really good results but then over time it's all about those signals. So you also fine-tune for large clients based upon their image libraries and then also it's essentially platform signals as well. Hey you taken down these 5,000 infringements. What are the commonalities between these 5,000? What are these signals? Oh, you've exported or you starred these X infringements. What was interesting about those and then using that to essentially use it as a feedback loop towards all right, we want to find more and more of those. Odds are they're going to be higher priority for you.

Yeah, that sounds like a lot of work. I can imagine like a tons of model building as well in your machine learning pipeline, right?

It's Yes, I think the main work for all this is just getting a feedback and to figure out what are the good signals.

Do those signals change over time or across different brands?

I think there are some golden signals that we've seen that work really well. They change across brands. Yes, they also change across platforms. Some platforms will have other different tells than others. But yeah, I think it's a huge part of really the hard work of this past year of development is just like, hey, how do we get better and how do we build this domain expertise into our evaluation?

Yeah. How do you balance this precision and recall consideration in these kind of situation? This is a common machine learning problem. Do you want to catch everything or just not flagging the legitimate sellers?

Yeah, it's a great question. I think for us really we start from the approach where we essentially pass it possible emphasize recall and then what we'll do is over time especially with brands that are very picky about what they enforce against get that rather down some it down into all right what is the top 20% of the cases that you prioritize so that's where we get to is we can cast this wide net we can surface on one run 30,000 results for you, but after that, how do we whittle it down so that I actually want to action these top 10%.

Right? And I think that's really where it gets down to in terms of things such as our image matching. Like this is something where we actually have essentially like bifurcated flows here where we have something specific like pixel pixel matching but we also have that fuzzy match as well because we know hey is it for example you're an entertainment client and you're making merchandise these designs get modified and they'll have different colors and things like that but at the same time it still is a copyright infringement. So that's where we also picked those up.

I guess if you hand it over to the human review, there's a little bit buffer of it's their discretion when it's the lines a little bit blurred to whether this is the real counterfeit or not. So they can make the call.

It's always that fine balance where it's like hey we want we want to make your job as easy as possible, right? We essentially want to serve these you on a platter and say hey these are good results but at the same time I also don't want to miss anything. So

Right, I guess the burden on the brand IP owner teams to really look into these when it becomes like a 5,000 a list of 5,000 it becomes really hard. They have to put a lot of manpower as well or these really or they just trust your platform. I believe what the top 10% from Podqi is going to be good to go.

I think this is it's combination of that. I think people get a lot more comfortable with time with using a platform. And so what we've seen is like now for some clients it's very quick right? They go into the platform and they can get everything they needed done within them an hour a week even others like to spend a long time on the platform which is awesome for us to see right but I think what helps there too is we also have a lot of automated flows on top of the results that we built for example if there is say we found one infringer one entity if I want to mark that entity as like a priority I can now go through and for any new results that come up if that entity shows up, I'll automatically like flag those and then you can go in and say, "Okay, we've seen these guys before. Let's clear them out."

When you're using anything related to large language models and unstructured data, is hallucination a problem at all? If so, are you trying to take measures to reduce it?

I think for us, hallucination, the good news is we do have a little bit of that buffer, right? It's like we serve these think of it almost as search results to you and if there is a bad result you can choose to ignore that and that's helpful. This is why we emphasize having that human in the loop as well, especially for that enforcement. But at the same time, hallucination is a problem, I think, especially when it comes to IP matching. Say I might have one image on my client's assets. All of a sudden, I hallucinate on how this image is being used on a listing for example. Now, all of a sudden, this is seen as match, but it's not actually a match. How do I be able to account for that? And I think this is something that we're still very much working towards and make sure that we essentially just have multiple layers of checks to say all right how do we make sure that it's not just one step of the pipeline that derails everything.

Yeah there's a lot of techniques around hallucination like making the extraction because you must have the extraction and a retrieval process right from your scraper results. So how to make it more grounded is a big question.

It's also for example there's different places of failure right say the scrape the JavaScript doesn't render correctly okay then I'm working with limited information or when I'm for example when I take the HTML I parse it or I summarize it what if it actually grabs raw information there what the hallucinates the, say, the ASIN that I get then all of a sudden this is a point where we have to pinpoint, all right this is where it went this is how it failed how do we come to that later part.

You told me last time that you're also hiring. Tell me more about it. What are you looking for and what makes someone a good fit for Podqi?

It's a good question. I would say the I'll get to the fit part first, which is we really want someone that loves to grind. And how I put it here is we have a very heavy in-person culture. Essentially, what we do when we work is we're just hanging out in a room. I out of the Bay Area. I think my two best friends are right there in the office with me as my co-founders and also with the rest of our plays as well. And I think the only reason that this works and this works well is because one, we're very aligned on we want to work hard, we want to get a lot done. And two, also there's no stress when it comes to working, right? Because it's in the end, hey, we're all kind of building together and we're I say having fun with it. Just like really experimenting, really chopping it up. And so I think that's like the one thing where we really want to prioritize having people that are very much aligned on that can work hard but also have fun when doing it because I think that's where that creativity aspect of things comes in and is really important for a startup.

How big is your team?

Right now we're a team of 11 full-time.

Sounds really fun. Technically, I'm really intrigued. I can imagine what you guys are doing and I think you're cracking a really tough problem but it seems to be something challenging but very interesting as

I think what really I'll speak for myself here like what really gets me excited is this is an awesome application of AI and it's in a space that has seen very little innovation over the years. Sometimes when I still run a lot of customer demos and just like for the fun of it or sometimes for sure that I'll do it. I love running these demos and we'll essentially run this. We'll take say a brand that we're demoing to. We'll take their publicly available assets. We'll do a search beforehand and we'll show them their results on the demo. And you're talking to the people and you're just immediately they see it and they're like, "Oh my god, I didn't even

No idea. Yeah.

I had no idea. And it blows their mind. And we're like, oh, and we catalog everything and this is what we do for matching. And they're like, we currently have paralegals doing this. I'm like, that's that's unbelievable. I cannot believe human no human should be doing this work. And I think that's the awesome part is like really a lot of this passion comes into hey like we see very tangible benefits. We make people very happy when they work with our platform. I think this is something that's it's hard to get in some other industries where people are like, hey, we're we're always being pitched new solutions. We're always being pitched new innovations. There's a thousand different dev tools these days. Whereas for IP, it's hey, like this is something where we want to do it and because this is something where it's an industry that's been going through.

What roles are you hiring right now?

Yeah, we're currently hiring a product focused engineer. So, this will be more full stack across the board. There. The idea would be essentially making them in charge of building out new features. They would be interfacing with customers, sitting on calls and hearing feedback there, and then working to build end to end. We're also hiring a backend engineer. Essentially, we're at the point now where, hey, we're doing really well with our current customer base, but if we 10x this in the next year, we're going to feel a lot of pain. So, how do we got this up? Essentially, how can we make our processing faster? How can we make things more distributed and get things essentially maybe even on the scaling front or on the scraping front that's been a huge part that we have to scale too. We're hiring there.

We're also hiring on outside of the technical end just more AEs, more customer and account managers as well just cuz now we have a lot of clients with all a lot of nuance needs. So it's now we should want essentially someone to be able to sit there and say hey I can be your point of contact. We're reaching that stage

For someone considering joining you guys what's some of the exciting technical challenges you're working on right now or potential they could be working on

Yeah so I think I'll say on two ends right which is one if you want to get into the nitty-gritty of things currently one of the things that we are working on is we're actually migrating are essentially like our task processing. So we're breaking down our tasks into even more granular chunks before say scraping task and then we have that evaluation flow and then we go from there. But there's a lot of bottlenecks here that we actually can get more efficient on. Say I have my scraping but at the same time I should start processing other scrapes other marketplaces when I have downtime between my scrapers. I have cooldown, right? For example, some marketplaces that I only need to ping or I have pretty heavy rate limits on. How do I make it as efficient as possible so that I can always keep my engine running and be able to be processing results, right?

How do I make full use of the rate limits that we have for LLMs while at the same time making sure that we're able to produce results and have quick responsiveness for the immediate targets that need to be action. For example, creep being let's just call like a random e-commerce marketplace. Those results refresh once a week, say every so often, but at the same time, say something like a Meta ads library that refreshes every day. Ads can go up the day off. They need to be taken down ASAP. How do we build in the ability to essentially prioritize some targets, prioritize others, and at different cadences for how we want to process those results?

For all my audience, if you're watching, just take some notes. This might be your interview question.

Yeah. We'll toss it in. Get the whiteboard out. A lot of us is I'm thinking about this at the moment. I'm like

Right at the moment

We can use some solicited help.

How do your customers discovering you right now?

So, this is something that's been really interesting for us, right? Where on the end, one end brand on the enterprise end, so I would say like on the larger end of spectrum is a very tight-knit industry, right? I think they're what's been doing us very well are essentially those industry conferences that we've been attending that we've been showing face getting a booth that and really putting the word out there and also a lot of targeted outreach towards I think brands that we have seen are having issues right or we know of when we can even do these sniff test searches and say all right I can tell that there's a lot of going around right and this is something that we want to get in front of you We can even pitch you and say, "Hey, these are problems that we've seen. Are you proactively can we get a chat in with someone your legal counsel?" But really, I think what we've seen, we just actually did a breakdown of our revenue sources. And now more than 70% of our revenue booked in the last 6 months has been from referrals.

And I think that's just been a huge thing for us, which is it's combination of the industry being tightknit and us getting the word around. And also too, it's just this is a problem that I think people are trying to find answers to however they can and they'll get into these groups or B toC like they'll have to subreddits where they'll chat or there'll be these industry connections and know I think for us it just means doubling down on doing a good job.

Let's say next year when we chat again and you may be facing a I don't know 10x 100x growth and you may be facing this growth outside of a referral stream of customers right how do you plan to go to the next level

Yeah first of all I really appreciate this you chatting with me I love these questions you had such good take on these but I think really the big thing is all right we're getting to the point now where we it used to be the new kid of the block and now we are established, right? People know of us, competitors know of us. How do we keep growing our brand awareness? How do we make sure that we are the top and the undeniable top brand protection platform out there, right? How do we keep making customers happy? And I think in the end, that reputation is what's going to carry us. That's what's going to bring in that inbound cycle. And that's where we want to get is just building on that, making sure that we're essentially keep doing things well, keep listening to our customers, and be able to establish ourselves as really this is the no-brainer option.

Yeah. I'm having a new closing question. I'm establishing another tradition on my podcast, and you're also the first one. Welcome. If you were to recommend one book to a 20-year-old, one to a 40-year-old, and one to a 60-year-old, what would they be and why?

Okay, I'll go with I'll go with the 20-year-old first, right? Honestly, I know there's a film about this, too, but this is one of my favorite books of the last two years was Boys in the Boat. And I know they came out with the film and all thing that but essentially brief synopsis it's about the 1940s Olympic rowing team based essentially being from University of Washington back when Seattle was very much a backwater kind of small mining bug in town and it just essentially details accounts of this you know college boys, athletes that came from nothing, right?

They came from families that were quite poor, quite down on their means and also during the midst of the Great Depression. And they essentially banded together, really gritted out, worked hard and became essentially established a rowing presence on the West Coast that had been dominated by the East Coast elites for the longest time. And it's just a amazing detail of just how they banded together and it really worked to build this against all odds. It's a classic underdog story. And I think about this sometimes when I'm sitting in the office of just a small crew that building out. It's just there's no replacement for hard work and there's really no replacement for just get the day. And that's been a huge part for us. And I feel like that was a really inspiring point that spoke to me. So I love that. I think when you're 20 years old is one of those where you read it and you're like h like I could be let's just do something big.

I love it.

And then I think a caveat on that is when I was just getting out of college I really started turning more towards non-fiction. I think that was like a big thing where I was like okay I want to read more non-fiction. I think this might be more applicable to my career. I want to get more learnings out of this.

And non-fiction seemed to be the thing at the time. I think I'm actually falling way back to fiction. I think in fact maybe more helpful to in my day-to-day now which is you read non-fiction but use account to build up that base and that knowledge but you read fiction or like creativity and I think one of my favorite fiction books of all time is the trilogy the Foundation Trilogy from Asimov. I think that's just one of those where it's amazing it's an amazing volume of books it really coming from decades ago and they're writing about things like artificial intelligence about robotics for the first time. It's a lot of things where you're like, hey, in a sense, people have always been thinking about artificial intelligence in a in a way. I think that was a quick thing I really need to reread it.

Fascinating. For people who wants to find you, whether they want to join your team or learn more or use your product, how do they reach out? I can share in our description as well.

So, reach out to me directly. I'm just Jesse, J-E-S-S-E, at podqi.com. I'm always on my email, so let me know directly. And yeah, if you're interested, hit me up. You can also visit our site at podqi.com and would I'd love to chat. I'd love to get a coffee if you're in the bay and really I'm just interested in finding and meeting interesting people.

Thank you so much. I really enjoy chatting with you and thank you also for the book recommendation as well.

Thank you, Angelina. Is amazing chatting and yeah, amazing questions. I'm a huge fan of new opening and new CL.