August 8, 2023
Rakesh:Putting AI in the name is very cheap at the end of the day. Just becausethey're getting more clicks doesn't mean people are purchasing more though.
Julian:Hey everyone. Thank you so much for joining the Behind Company Lines podcast.Today we have Rakesh Yadav, CEO and founder of Aidaptive. Aidaptive's Missionis to empower any size e-commerce or hospitality brand with the power ofenterprise grade machine learning. Rakesh is so exciting to have you on theshow, not only because of what you're doing with Aidaptive and obviously.
Ai, everything.Everyone's talking about it. So, we'll, we'll dive into some interestingtopics, but you, you've been building machine learning for such a long time andyou've had such a fruitful career with that type of technology. I'm so curiousto see how it's evolved, where it's currently, and almost the speed at whichyou're seeing, new models being built.
But before we getinto all that good stuff, please tell the audience, what you were doing beforeyou started the company.
Rakesh:Yeah, no, I I did a master undergrad India Master's in us and then joinedGoogle as first job. But even during undergrad days, I always wanted to buildup a company While in Google I wanted to build up the skill sets so that isneeded.
So to increase theprobability of success of the company. So that's why I think of my life in likethree different phases. One is like first like first 15 years where I built up.I mean the work life where first 15 years I've built up the skill sets. Second15 years, I'm gonna leverage all of that.
And then third, 15years I'll See how it can contribute to like, back to society around, like someharder problems like energy healthcare and things like this. And Google yeah.Thankfully I was lucky enough to be called as a textbook entrepreneur. So whilewhile I was still in Google and all the cushy job benefits and all that, Iwould still call was getting called in after like first few years once I becamesenior to, yeah, start zero to one projects.
So thankfully wehad created like 14 projects in my 14 years and Few of them, like eight of themreached a billion plus a rrr, and seven of them I killed without because Ididn't see a within six to nine months. So that taught me a lot more. I meanbasically I was building up my startup skills like market.
Yeah. Building upmarket, how to come up with a strategy, 3, 5, 10 year plan how to add this.
And, and were youkind of coming up with the ideas on, how to, what, how to leverage machine learningor some piece of technology, or was it, what in particular were you trying toessentially bring to market?
Rakesh:Yeah, so one of my most recent projects before I quit was ads new machinelearning pro platform. So where yeah. The premise was, prior to that I wasleading a payments machine learning platform. So basically the entire moneymovement that people are purchasing in Google, like from Android apps to buyingGoogle home devices to, phones, to YouTube subscriptions, to everything cloudbuying cloud credits or whatever.
All of that wentthrough the payment system so that we'll identify fraud. And when like one,yeah, one 50 billion is flowing through your system, like five, three, 4% fraudis actually a lot more significant. So yeah, thankfully we did great, but thenI moved to ads and over there the. Before setting up the ads ML platform, themain problem that I was handed off was how to increase the innovation cycle inthe ML world of the ML application.
Google Ads uses MLall over the place. Yeah. 'cause at that scale you have to use machine learningand for idea to reach production, it used to take like six to nine months andthat was like too slow for our guest. So we set up a new team. We startedbuilding out new infrastructure with the North Star being let's us see if wecan get an idea to production in like a day.
Julian: Iwas gonna ask, when you mean, when you talk about the innovation cycle,describe that a little bit more. What, what does that mean in, in the languageof machine learning?
Rakesh:Oh, as in like, let's say we have a new idea, which is a let's build out a newmodel to predict let's say Google shopping, right?
Like Googleshopping ads. Yeah. When you're looking at like, all this inventory and we wantto build a model to predict when an inventory will go out of stock, partlybecause there's like, weeks of delays of like merchants sending us data toGoogle and all of that. So we want to build out a model to make it moreaccurate.
Yeah. So, so that.So that is the business statement. But from there on, to actually build out amodel to train all this data at Google Ads scale, which is like, terabytes andpetabytes of data. Yeah. And then have a model trained, evaluated, then putbehind experiment and ramped up in production. So that we can materialize thatresults.
That end to endrequires so many challenges. Yeah. That it is like six to nine months. I wantedto reduce that, like less than a day.
And how do youstart going about, doing that? Is it building more sophisticated models? Is itbreaking down its components and, and kind of con is it going Yeah.
Through each stepto figure out how you can optimize it or connect or even skip steps. Where doyou even start with thinking about how to decrease? That, I mean, that timewindow decrease that amount of time from six to nine months to a day, I mean,that's incredibly fast. Or in an instance it's incredibly fast.
Where do you goabout thinking about shortening that timeline?
Rakesh:Yeah, so first thing you do is first thing is give a control plane to all modelengineers. So one of the biggest challenges is model engineers are typicallynot like in French engineers. Definitely not at like Google ad scale. And thenin engineers typically are not like model engineers who don't understand thelatest cutting edge, modeling techniques.
So even if amodeling technique would work in a offline where with like of training systemsat even Google, there were like plenty of training systems. TensorFlow beingthe most popular one. A bunch of internal ones. We probably didn't expose inopen source yet. And we consolidated everything in TensorFlow.
But with my systemsand techniques, even if the model that performs and provides like decent lift,getting it to production and putting it in like production was like additionalchallenge in addition to like, yeah, the modeling challenge. So first thing youdo is put out a control plane where you hide all of these training systems andabstractions out.
So that way aresearch ML researcher, if they are picking up a new cutting edge technology,they have a very modular way of bringing their technology with all ads usecases. Yeah. But then the ads products actually are operating on that productplane, like control plane. And they don't really, they just pick and choosewhich model or techniques they want to do.
Yeah. So that isthe infra control. And then model ML research coming together in like oneunified platform. And then there's like the textbook big data challenges oflike Google ads, trying to process like tens of terabytes of petabytes andexabytes of data for every Yeah. And then running it all every day in ascalable and monitoring fashion.
So those were thethree big steps.
Julian:It's fascinating thinking about, where you've gone to with, with Aidaptive andbecause I've learning a little bit more about the company and kind ofunderstanding consumer data is a huge way that, I, I think you, you'veimplemented this technology and.
Was traditionally,that timeline you described earlier, was that what a lot of brands ore-commerce or hospitality companies was, was that what they were facing as thismodel was continuously built and build this profile? Were they seeing that typeof I guess timeline, with the technology that was existing?
Rakesh:Yeah, more than that timeline also just to like, many of the folks that sinceML is such an asset and constantly evolving field, Many of those folks don'thave the AML talent in house also. And even if they did, yeah. They didn't knowhow to like evaluate them and what to really bet against.
Mm-hmm. So that'sthe additional challenge in addition to like, building out something andmaintaining it and constantly running. So that is the infra challenge. And thenthere is like the business challenge where market dynamics changes every day,like Covid taught us, right? Like every business had to adopt very quicklyduring Covid when Covid hit.
That would meangathering new set of data, training your existing model. Again, evaluating thatthis is performing well and then swapping it out in production and and then alsopicking the right technique among like marriage of different techniques becauseyeah, like. 20,000 ml algorithms and models and things are existing over, andwhich one applies for the business.
So all of thatactually increases the timeline beyond like six to nine months. Even if theycan get it done, like keeping it up and running is also a constant effort.
Julian: Iwas thinking, one thing you just said is, is applying ML to the right type ofbusiness model, the right learning model or the right, that machine learningmodel for the business.
How, how do. How doyou do that? Like, and how, what are the different models that are available?Is there an index that people use? Are certain professionals, kind of preprefer others? I, I've learned a lot about this. There's this whole researchcomponent that goes into machine learning and then it gets adopted.
Where is that indexthat, are we able to access or just machine learning professionals who goesabout and match, does that matchmaking process.
Rakesh:Yeah. So if someone is like, let's say while we are in Google, we would look atall different techniques in which we'll apply to d given use cases because youhave to understand all of these different techniques that's part of yourresearch job.
But as part ofAidaptive, and if someone is Aidaptive client, we really all the research andeverything else gets happens on the Aidaptive side versus our customers getlike very stable product APIs like recommendation search. And then depending onlike individual customers, data shape and patterns among like, like 20 models,the best one gets picked by through like a automatic experiment system.
So we basically,like, no, humans can pick something. We just let the best one get pickedautomatically for a, yeah.
Julian:It's, wow. So not only have you automated that process and made sure that'scompletely streamlined, but then now this whole, kind of almost immediateconsumer profile that you're building, how is that accomplished? Because,because no one has done anything yet. If I go to a new e-commerce website, if Igo new to a hotel, I haven't necessarily given the system anything to referenceprior.
So how is that madepossible?
Rakesh:Yeah, while you say you don't, you haven't given anything technically, we stillhave like, device types, IP address your geolocation. Yeah. And then as youkeep clicking, browsing through, like hotels are products. Each of those areactually building profile.
So what we havedone is we like the recommendations and such is our own widget as well. So wecan actually collect all the data that we want for our training system so wedon't depend on cookie or anything, third party, which is, which gives us likenext level of independence. And after that, then we do really cohorts, even ifwe don't know enough, we just do a cohorts of, like, this customer is, like,for example, you are from New York right now, who used to live in LA in thepast because your previous IP was from la.
Sure. So that, andthen you have like high L T V versus rockish as like low l, t, v, like lowpurchasing power. Yeah. And you, we both prefer black T-shirts o over timebased on, as you click through products, we can build out that you're onlylooking at like male t-shirts. So there, there's, there's a different basespace model that gets built out.
And then based onthose cohorts, we identify in the known purchase, known customer cohorts.Somebody who like the cohort, which is similar to this. And then we smear thosekind of recommendations over to this unknown cohorts. Yeah, and that's where Ithink, that's where the holy grail is at the end of the day.
So it's almost like a comparable when somebodycomes in and you come with match them. And, and what is this? Obviously itallows companies to offer the more specific, accommodations, the hospitalitybut what does it mean in, in regards to the additional offerings that companiesare able to then do it?
Because I feel likea lot of the promotions or a lot of these. Things that companies do, they'renot necessarily landing with the consumer as much as they used to. It. It seemslike the consumer needs have shifted, but a lot of companies haven't shiftedtheir offerings.
Rakesh:Yeah. That's why we actually trade, because the, the consumer buying behaviorchanges, inventory changes, like the rest of the world is constantly adapting.
So that's why wekeep changing. We keep adapting our models on daily slash weekly basis based onthe cons, based on how much data a customer has and hence the name Aidaptive.But the key point though is for a given given consumer pattern, at the end ofthe day, we really can pinpoint we can really adapt very quickly, like matterof days slash weeks to any changing behavior.
And then what kindof offerings do you see and anything that surprised you in how they've utilizedthe data and the information that, that you've, collected Oh, yeah. Andpresented.
Rakesh:Yeah. Yeah. So in addition to our like the textbook products, which is like,recommendations, search and audiences.
In addition tothat, we also have like foundational things like how I saying like, Purchasingpower L T V for every user of our clients. We compute all of that, and then wegive that. So one customer was very creative in the sense that they used our LT V for every customer that we generate, and then they actually personalizedthe buyer experience or even customer tech support.
What I mean by thatis, like I R L D V will go into separate phone support, which will have likeless than. Five, second wait time and that the mid and lower LT will go in theother pool so that way the higher LTV gets like, premium support. So that waslike very good, innovative use case of our foundational tech, which we didn'tthink about.
Julian:Yeah. No, I, I, I, I wouldn't have thought about that either, but that makes somuch sense, just getting to the more, I guess, more ready buyers right.Immediately and supporting them. I mean, I, I think that any company would say,yeah, that's a pretty good idea. Yeah. But it's so fascinating to think aboutthe, like, continuous ways that, that companies can still adopt yourtechnology.
Anything that yousay maybe hasn't been adopted yet that, that you'll think, is on thehorizon.
Rakesh:Yeah. So for example, in ways it'll be used. Yeah. So for example, search is avery commonly used term, but when you are searching for a shirt on an apparelsite versus when I'm searching for a shirt on an apparel site, it should behighly personalized.
It should not bejust like, text based, like matches on the entire queue. It should have like,my intent, my persona, like. Not just my persona, but also like if I'm aboutto, if, if I'm purchasing for a business trip versus if I'm purchasing about togo for a hike or a vacation on a beach, that is the intent as well.
So there's ultimatelevel of personalizing that and that search results. I think we are in the veryearly phase of that and there's like lot of leeway where we can really improvethem and of course the revenue of the customer, but more importantly than theirpurchasing experience.
And it seems like alot of people, I, it's been said on the show before, but you know, everybody'skind of craving this up to date kind of consumer experience, no matter, whatthe product is. And, and now it's almost standardized to have that kind ofreally facilitate experience. And if you're not, you're not competitive.
One thing thatreally kinda came to mind thinking about the way the technology's used, Whyhospitality, you know what, why hospitality and, and why is it like certainmodels work really well for hospitality. So you wanted to double down in thatdirection. I was trying to, figure it out and on my own, but I'm curious, whyhospitality that, why do you think this industry has such an ability to,benefit from this technology?
Rakesh:Yeah, so we really just sell the same three products on both verticals. Andthen we'll keep adding verticals just that the entire modeling and whatnot, andsurvey and the data schema and knowledge graph behind it. To understand changescompletely on a power vertical, but everybody needs a recommendation.
If they're sellingsomething, a recommendation engine, everybody needs a search, which is highlypersonalized. Everybody wants to understand their customer base with like, allof this different colors like taste, purchasing, power, propensity to purchase,all of that. So by design I've always wanted to pick one more vertical becausein the background we are really building again that end-to-end ML platform.
Like so far our ownmodel and infra its users. While we are selling product, we are really buildingthat end-to-end ML platform, which we can open it up to developers tomorrow.And to keep it generic and to keep it honest. That's why we always need morethan one vertical, otherwise it'll end up becoming like a e-comm platform.
And then onehospitality specifically was like, through just my network. One of the firstone of the first customers was in a hospitality industry. So I was like, Hey,do you, are you interested in this? And they're like, yeah, why not? So like,sounds good. Technically, we recently got into like boating business as well.
We are in the earlysafe phase where again, network, like word of mouth, they came to us and like,yeah, of course if you need recommendations, sure. Search. Yeah.
it's, it'sinteresting 'cause it seems like Aidaptive is really just a, a, a reallyvaluable tool in the pursuit of whatever these companies are trying to do,deliver, value to their customers.
And I, and that's areally unique position because, it's, it's almost like, as you progress and asyou, you're incentivized and consumers are both incentivized by the same thing.It's. It's not very common with when companies can find that unique businessmodel. So you know, now that you've been, building Aidaptive, you're workingwith a few different, industries, verticals, focusing on some, describe notonly the traction you've seen thus far, but also where do you go from here?
What direction doesAidaptive go?
Rakesh:Yeah, so 3, 5, 10 year plan. So three year plan was to build out these productsand then extract have a G T M motion going with like a good product market fit.Mm-hmm. Across at least one, two verticals. So that's what we are focused on.We are in like a, we just finished, we are two year old company, so this is thethird year five year is like a scale all of that up so that we can keep addingto the entire e-comm funnel.
So can we keep addingto the suite? Mm-hmm. Yeah. And then also we, the other dimension to scale upis not just us north America, but go eu, apac, all of those. Yeah. The 10 yearplan though is once these verticals are up and running and we have like aplatform, then actually open that up and then we become one of the hopefullyeasiest to use well contained end-to-end ML platform in the world that willallow the developers to leverage it.
So that's a 10 yearplan.
it's amazing tothink about it becoming, leverageable by developers because I feel like a lotof there just, it opens up capabilities for companies to adopt the technology.And does that mean. These new will, will, these developers have to learnanything new to be able to, use these types of pieces of technology.
If I was a machinelearning developer now, and I'm looking at this, I'm like, what am I gonna haveto know to be able to use a platform like this or other tools to be able tomake myself not only competitive, but you know, uplift my company or boostwhatever piece of product technology, or bring something to the table, what doI have to learn?
What do I have toknow?
Rakesh:Yeah, so the control plane where you're controlling the training as well asproduction model in production or like the inference, a p i, that control planewill be the only thing new. But in the background, for example, for trainer, westill we still support like TensorFlow and PyTorch and then for like, having itin sewing, we again support like TensorFlow light and, serving and we do havesome craft mining techniques as well.
So all of this willbe, are like standard things, which are well known for researchers. Just thatthere'll be a control plane where all of this is like, in a well controlled,hopefully easy to use ux. We'll see how we that. And then yeah, also, we'llhandhold so that the places where people shoot themselves in the foot, like forexample training, serving sku, like all of those will be inbuilt already.
So there'll be likea lot of checks and balances in the in place.
Julian:What do you view as some of the biggest risks to Aidaptive today?
Rakesh:Yeah, so one of the I mean there are plenty, it's a two year old startup. But Imean, the biggest one of course is like my, like a bunch of us are ex Googlersand my buddies.
But if a bunch ofus leave, that's probably the biggest one. Hopefully that never happens. Butthe second second biggest is in terms of just business side. I think in thee-comm world we have been talking about personalization for like two, threedecades now. And a lot of these sellers have been burned by vendors who talk AIml, but they don't really do like real AI ml, so the results are likeunderwhelming.
So they like,because Aidaptive doesn't have a big brand presence, they actually are a lotmore cautious to, like you use someone like us. So that just increases, thatjust gives us a little bit of a headwind. Thankfully we are working throughthat by just saying that, Hey, how about you use us for free if you don'tdeliver value?
Uh uh. No harmdone. Just getting, we want to build like 10 year old, like 10 year longrelationships at the end of the day. Yeah. Not like, just make money, moveon.
Julian:If everything goes well, what do you view Aidaptive, it becoming, what does thecompany become long term? I know you mentioned a 10 year plan, but Yeah, yeah,yeah.
Rakesh:At at least Aidaptive. So the platform play is what we use the Jarvis ml, whichis the original name. So that will, so that, so that's why you'll see AidaptivePowered by Jarvis ml. So that will get opened up later. But on the Aidaptiveside, ideally we have like very well contained intelligent products, which arevery easy to launch and integrate.
Yeah. So that waypeople can leverage the power of ai. Basically AI is currently limited to like,top 1% of the companies and rest of the folks are not like leveraging it wellenough. So the, the main premise is we want to democratize the access tomachine learning and leverage its power to generate revenue for rest of thebusiness owners.
Yeah, while, whilehiding all the complexities of the latest buzzwords like Jen ai, how does ithelp my business? Like, I don't think many people can answer that for their fortheir own business. But thankfully, Aidaptive hopefully will have like suite ofproducts, which will answer that. Like for e-comm, gen AI could be used togenerate titles and descriptions and have better understanding of like theimages that you're using and give you feedback that this image performs better.
This type ofdescriptions performs better. Uh mm-hmm. They understand the consumer behaviora lot better, so that way you can invest in this, that your com consumerpatterns have shifted. Are your marketing strategy shifting? So at the end ofthe day, we want to be like the intelligent partner to help the business growrevenue.
Yeah. Andconstantly adapt to changing world dynamics.
Julian:Yeah. This next section,
I love this next section I call my founder, Fa Q. So I'm gonna hit you with some rapid fire questions and we'll see where weget, so, yeah. Yeah. First question I always like to open it up with is, what'sparticularly hard about your job day to day?
Rakesh:Lot of it is actually the context switches, which is from like legal meeting tolike operations. We are in three countries to like, then some model questionsabout some model underperforming then to like infrastructure, then some fireabout like, tax or accounting. I think that context switches is just it's notfor everybody.
Yeah, and then I, Ikept, I made it probably a little more complicated by keeping it in threecountries, so,
Julian:Yeah. Yeah. Yeah,
you mentioned The,the idea behind real AI and describe that because I was, I was actuallylistening to one of the podcasts that's on your website and, and they weretalking about this concept between real and branded.
And I hadn't reallyconsidered what that meant and also what it meant for me, say if I was acompany purchasing a product of either way, please describe it for ouraudience.
Rakesh:Yeah, definitely. So, Like putting AI in the name is very cheap at the end ofthe day, but then actually identifying among the married list of technologies,which one fits the business.
So, for example,one technique which fits the business on apparel might not be the same for likehospitality. Of course there are some things that which will translate verywell. For example, our knowledge graph where we build out a taxonomy. Bylooking at title, description, and images, that one will translate whatever.
There's title,description and images. So there's like commonality, but then there is stillsub-vertical specific things which require like plain vanilla modeling doesn'twork, which require like multiple layers of, so at the end of the day, youreally need like, ensemble of modeling techniques to work in one cohesivefashion.
And then optimizingfor like very clear goals like Do you want to increase the revenue per visitor,average order value c t r. So things like those. So that's the goal. But thenwhich technique works best? And not just that you keep adding more techniques,right? And because you, you never know which works as well.
So then runningthat and a very automated fashion where the be the winner wins. So the businessgets the best technique all the time and then are constantly adapting on dailybasis. All of this is a lot more harder for like a typical brand who has notdone like, let's say AI or just doing rules because any rules.
Mm-hmm. Or if themodel is trained and not trained again they will not be adapting to changingworld. Yeah. So that's the key difference between ai. I mean,
You mentioned theconcept about measuring performance of, whether it's new campaigns and what it,or, or certain images and things like that.
When you, whatmeasurements are people using to identify that? Is it, clicks? Is it how muchyou view something or how long you view something, how much you view similarthings? How much of what goes into these measurements that people don't really,aren't really aware of?
Rakesh:Yeah. When, when you're a business owner who are selling something likecreating a D T C brand you're really more worried about lot more on the higherorder business metrics, which is like total share, total revenue average ordervalue revenue per visitor.
And then whereinthat funnel the money is getting, like what is the funnel flow rate like we, Ihave paid so much to get so much in ads to get people on my site. How many ofthem are actually converting into. Engagement to like real purchases, to likerepeat purchases, right? Like the loyalty, which is the last part of thefunnel.
So those are thebig higher order, big business metrics we measure. Then the second higherorder, which is like, engagement, like C T R, click through rate, all of that.Just because they're getting more clicks doesn't mean people are purchasingmore though. So that's,
Julian:Yeah. Is, is Aidaptive able to kind of help with, I feel like a, a lot ofe-commerce brands or direct to consumer brands?
I hear the biggestchallenge is, is when they have, a customer who has something in their cart whowants to then check out. It's that process right in that finish line. That isprobably the most challenging for a lot of them to figure out, do you see Aidaptiveas a healthy impact? Or I guess, understand and extract reasons why thatdoesn't occur.
Rakesh:Yeah, so our recommendation and audiences where we generate like the marketingemail list every day, they both take that into account. For example, whensomeone is in the cart, our sellers can embed our recs. So we can also showlike products you may also like, like accessories or some other offers inaddition to that.
So that increasesthe the cart value. And then second is if someone didn't purchase that is a bighint for our audiences. So we can just reach out in the email campaigns to themagain, like, Hey, you were thinking about this. Yeah. Have you, if you didn'tlike this one, have you considered these other products?
So same Revix, weput it in that email as well. And again, it's like one more touch point tobring them back into the purchase flow.
And in regards to Ifeel like a lot of people when they think about from, obviously from a brandperspective and a com, a company perspective, you want to deliver exactly whatyour company or your, your consumer's looking for.
Excuse me. But froma consumer's perspective, sometimes you get a little bit skeptical or I guess,wary that you're not necessarily uncovering or discovering new things that are,branches of what you originally liked. Then leading to something else. Doesthat ever come up in, in a lot of these conversations, are companies worriedabout that, that, they're, they're kind of gating consumers?
Are we seeing ashift in that? I, or am I just the only skeptic that, that thinks thatway?
Rakesh:No, no, no. I, for example, any company I, if I'm, if I have like, let's say2000 successful products SKUs, and if I'm launching 10 new this month, How willI get people to know about them? Because if I keep showing those 2000, likemost successful ones, then there, so there's like, there's constant, like youwant to have people discover these products as well.
And that's where abunch of our search algorithms actually wait. New products a little bit higher,so that way, and then we also have like a categorization system where you said,mm-hmm. When you said browse, so you can browse like, oh, trunks, but thenbeach wear. Yeah. And then you, you narrow down with like contextualinformation and then it gets ranked according to like your preferences.
So tho those, allof those, all of those categories are, can be autogenerated by ml. So we don'teven need, like throughout taxonomy and we don't even need like, someone,someone human to write all of those down.
Julian:Right, right. As, as Web three kind of comes into play with a lot of data justchanging, right.
People being ableto opt in if they're, in some kind of business or website or whatever, andtransacting at a marketplace, how does that change or does that impact.Aidaptive in a positive way. I can only see it impacting it in a positive waybecause you only know more about the people you know more about.
Right. I mean, itit, but how does it, I guess, grow in terms of this whole idea about brands nowpartnering together using similar customers and things like that? How do yousee Web three and that? That's probably way down the line, but do you thinkabout that often?
Rakesh:Oh yeah, let's, let's, let's say like, best case scenario, everybody's usingAidaptive, right?
So there is somethingto be said about like, while anonymized, like having a consortium where allsellers are helping each other out, and then we can track the user journey, notjust on my D T C, but across all of our DTC sites. And then we can give likevery highly personalized experience throughout the entire consortium ofsellers.
So that's like veryfar out, but yeah. That will be like very, very interesting. Once we reachscale is one of the Yeah. Levers we have to put. Definitely. Yeah.Engaging.
One thing I wasthinking about with, with this, this relationship between brands and consumersI'm, I'm curious to hear your take, what, what's something that you've seen orwhy, why do you feel that.
I feel like brandskinda led the direction of a lot of tastes, culture, things like that. But nowyou see almost inverse and they're almost fighting for attention. Is it, is itjust a mere fact that they're not leveraging the right tools? Or is there somekind of, more abstract decision making or, considerations that people aren'tmaking?
What do you see as,as where the disconnect lies or what, I guess, what's your opinion?
Rakesh:Yeah, so I mean, So the main reason is the entire consumer relationship hasbeen typically managed by folks like Amazon and all because they have investedso heavily, at least on the purchasing site, that all sellers, it's very hardfor sellers to create their own d t c like direct brand.
And then on on a,on Amazon side, I don't think Amazon gives them the customer relationship orwhatnot. They have to go through the Amazon portal. Which is by design. It's agood business model for Amazon. Yeah. That's why they don't really like,sellers don't really own the customers at the end of the day.
Mm-hmm. That's themain reason. Like a lot of traffic has shifted on, at least on the e-comm side.A lot of traffic has shifted from like search also actually just Amazon search.And that has become one of the key factors that it's harder to build a brand.
One obviousquestion I had to ask you is how did you stay at Google for 14 years?
That's a longtime.
Rakesh:I, when you are having fun, you'll be surprised how fast time flies. So, if, ifso let's say if I, I get involved in like unsuccessful product. So that's likesix to nine months and then I kill it. But those six to my nine months are veryintense. And that flies by like, very fast.
Then there's likeone month of thinking whatnot about like, new problem. And then let's say onthe successful product side, it's like two and a half to three years, sometimesfour until it becomes stable. And I'm, I get called for like something else.But even those, like, it's like the textbook startup cycle, right?
Like the first overone year, you're like, ah, barely working. Then second one year you're like,okay, something is working. And then third four, you scale it up, lock it downto Google scale. So all of this, what you're having really fun is just goes byvery fast.
Julian:Yeah, yeah, yeah. No, that makes sense. That makes sense.
I, I always like toask this question,
whether it wasearly in your career or now, what books or people have impacted you the most?
Rakesh:I. Yeah. So I mean, books wise there are like plenty of books. Like for examplelet's see, which like, for example, like I have a Dear Founder book, if you'relike a startup founder, that's a good one.
Mm-hmm. Zero to oneis amazing. Mm-hmm. This will make you small, I guess.
Julian:Yeah. I, I guess maybe not, maybe I'll, maybe I'll rephrase the question. Maybea non founder book. What's a, what's a book that impacts you as a founder butisn't necessarily about business or startups?
Rakesh:Yeah, so, one of like few of my favorites are like, think and Grow Rich, whichis like how you should really build up the skill sets and things that areneeded.
Second is likegrit. Like a lot of it is just Just gring it out there, uh uh, and notquitting. That is like a very good, like as a parent, that's a good book. As aself-development. That's a good book. As a startup founder, that's a good book.Like that's as a student, that's a great book. Uh uh, then there is a, I mean,I have written on my blog, I've written like 42 rules to live, live by.
That's actually,that's my way of like living, like that's, that has been assembled through abunch of books as well. Yeah. And then of course I love autobiographies likethe Sam Dalton Walmart guy. Yeah. So that work autobiography, that biography isamazing. Benjamin Franklin's is amazing. Bram Lincoln, of course.
Yeah. Yeah, Tesla.Tesla is great. Yeah. So those are great. Then of course, they're like businesslike, individuals who really give like, like I derive my I derive my drive fromlike typically sports persons or like, military folks. So there are like plentyof, uh uh, yeah. Like, for example, start with why Simon Sinek, he has likeamazing research from, military folks.
Like what drivesthem? Yeah, like Navy Seals, how they're training and how they watch out foreach other. That's like amazing. I. Yeah. Again, like life, not just likefond.
Julian:Yeah. I like the, I like the mental models. That, that models.
Rakesh:Yeah. Oh. I have like too many, like in my office, I can start looking at thebooks that are there and there's like too many right now.
Yeah. It says thiswill make you smarter. Like this is a good book.
Julian:Yeah. Okay. All right. I like that. I like that. Yeah. Yeah. Well, RO, I knowwe're coming to the end of the show here and I always like to ask, is there anyquestion I didn't ask you that I should have? Anything that we left on thetable here today?
Rakesh:Oh, no, no, no. This was great and also I didn't have much agenda. Thanks a lotfor having me. Yeah, there's always plenty of questions that can be asked, butthat's like 39 years of life. It, it can be summarized in nine minutes, butthanks a lot for hosting. This is great.
Julian:Thank you. Appreciate it. Yeah, thank you so much, rake.
So it's such apleasure not only learning your early experience and kind of the trials of, ofkind of startup within a bigger startup at Google, which is such an interestingexperience, but also. How you've created this incredible progress with, machinelearning and, and the modeling and really helping companies actually Deliver tothe consumers, really what they're looking for.
And I'm reallyexcited to see the brand expand and, and become bigger and, and impact me evenwhen I under the hood. And, and maybe I don't even recognize it, but lastlittle bit is
where can we findyou? Where can we find that blog? Give us where your, your plugin, yourLinkedIns, your medium.
I don't know where,where we can go your sub stack. Where in particular can we find, be a fan?
Rakesh:So my personal blog is on yadavrakesh.com. But you can find my profile onAidaptive.com/about and actually not just me, my whole company is filled withlike interesting candidates and it's filled with their profile.
Julian:Yeah. Amazing. Rakesh, such a pleasure. Thank you so much for joining us onBehind, Company Lines today.
Rakesh:Thank you. Thank you for hosting us. Thank you.