riverside_rui_raw-audio_floating_questions_0029
===

Rui: [00:00:00] Welcome to Floating Questions, the podcast where curiosity leads, we follow and stories unfold. My name's Ray, simply asking questions, shall we begin?

Rui: Hi, Ashish. I just came across your, book online, fight fraud with machine learning, because I'm in the similar space and I'm just looking for, other people's experience and insights. And that's why I reached out and thank you so much for accepting the invite to be on the podcast.

Rui: How are you today?

Ashish: Hey, Ray, I'm good. 

Ashish: thank you. Uh, yeah, I'm glad I had this opportunity with Manning. the book should be out in the next couple of months because we are in the process of finalizing the last chapter and doing the final revisions and it is [00:01:00] already available in the Manning's early access program. So anyone who wants to read the chapters one till the penultimate chapter can already do so.

Rui: Okay, I was like, well, it seems like there are more chapters. And those chapters are the ones that I'm actually interested in, but I couldn't really access it. Now I understand why. let's, let's just start with a self introduction. Um, would you like to talk a little bit about your background?

Ashish: Yeah, I might go for too long. So feel free to stop me and ask questions. my name is Ashish Ranjan Jha. I am originally from India, born and brought up different parts. I studied in one of these IITs, considered to be the premier universities in India, 

Ashish: while contemplating my options, to do master's, 

Ashish: I found some really good ranked universities in Switzerland. one of them was EPFL. that is where I went to do my master's in computer science. there was a relatively new, and. pretty famous at this point course at [00:02:00] Coursera from Andrew Yang on machine learning. I think this was the first course on that platform.

Ashish: I did that while being in Bangalore and, machine learning kind of, you know, grew on me. I was like, let's do this as my specialization during my masters. So at EPFL, I just. went full in, I did a master's thesis with Sony in Stuttgart in Germany, on some audio data sets.

Ashish: when my master's journey ends. And having worked at Oracle. I had realized I do not really want to work in a big corporation, because I wanted to have more end to end responsibilities. and I 

Ashish: went for a startup in AI space. in Belgium, in this place called Antwerp. It's a pretty nice city, very underrated. I highly encourage people to go and check out. I worked there, lived there, for three years, at Sentience, the startup in Belgium.

Ashish: And that's, that was my actual real [00:03:00] experience of learning all the different kinds of roles that you need to do at a startup, as a data scientist, as a machine learning engineer, as a product manager. Team lead, et cetera. this was already five, six years of my journey into Europe. And at this point I lived across three linguistic regions, which is the French part of Switzerland, German.

Ashish: part of Germany and, the Dutch speaking part of Belgium. I tried learning all of these languages, uh, learned all of them at a basic level. And I was increasingly feeling I should maybe go to a country that speaks English so I can better, find myself, placed in the society.

Ashish: Uh, UK hence was in my radar and I, applied for a few jobs. I got, an offer from Tractable, an insurtech startup in London. It's now a unicorn, I believe. So I worked there for a year. At that point, I decided I should start up on my own. [00:04:00] I applied for Antler. this incubator program, they accepted me, that triggered me to apply for this talent visa because you cannot start a company as an immigrant if you are just on an employee visa.

Ashish: And so I applied for it. I got the visa, but it took a little while. By that time Antler's cohort had started and so I lost the train. And as destiny would have it, that is when I applied for my, then four year stint at Revolut. and I would count my experience at Revolut thus far as being the most, important so far, in my career 

Ashish: I joined Revolut in late 2019, back when it was recently a unicorn. and I left 2024 Feb. when it was valued at over 40 billion. So I saw that 40X almost growth in the company, it was a great experience, to be in the company for that duration. We built a KYC system for the [00:05:00] Revolut app essentially from scratch.

Ashish: and that is where, having led a few fin crime related teams, uh, building the product. All of that experience, I felt I had to materialize in the form of this book, and that's where the idea for this book came about. and one thing leading to another, I finally came back to the idea of starting my own company.

Ashish: so here I am, three months into building Native. Um, which is funnily enough, uh, based around localization. So my exposure to different countries speaking different languages. Kind of motivated me intrinsically that there should be something in this world that enables people to access content pretty easily, be it of different languages, 

Rui: Fascinating. Uh, well, first of all, You have just given so far the longest self introduction in my podcast, which is great because I love how people are just like, without even me, prompting giving me the full, quick history of their life so that I [00:06:00] can start pulling the threads about so many interesting points, uh, I guess, maybe let me start with You know, Asians, emphasize a lot on education. And so elite institutions and education background has always been a theme. It's very apparent in your journey too. Would you say that It's really part of your identity.

Rui: How has your point of view around that has evolved over time?

Ashish: absolutely, the general culture I would say in those countries is around like, hey, get the degree from the most elite university possible. I would say in general, this is becoming less and less true. From whatever I am perceiving, listening to other podcasts, talking to other people, uh, also looking at younger generation of India at least.

Ashish: people are becoming more confident and not so much reliant on degrees. When I was that age, of doing the bachelors and the masters, when you come from that [00:07:00] culture, you usually tend to abide by the rules.

Ashish: And so I played by the rules. but I would say looking back, I haven't really leveraged so much on my degrees as much. many people, actually went on to the PhD, even if they were not sure about the academia, because.

Ashish: They just had that, title, the degree, the big university name attached, to it. But I took a risk that I will go to a no name startup, uh, to a no name land to explore things on my own. And maybe that's not for the best. but I think I have transitioned from attaching myself too much to the names of these universities, to just actually being useful in what I do in my job.

Ashish: I think today's breed is definitely a bit different. in India, for example, this is a recent shift, a lot of 20 to 25 year olds are actually doing their own startups by that age. When I was graduating 10 years ago, this was absolutely not [00:08:00] the case. People were just looking at which job can I go to?

Ashish: What is the biggest company I can get an offer from? Starting their own company was really like a long shot, a really huge risk. So what was something maybe for the elites in a decade ago is becoming more of a, um, mass, thing right now. and in general, yeah, education, I think is going through its own identity crisis in many ways.

Ashish: It should be questioned what we study. So as I told you, I did my entire four years degree in IIT in electrical engineering. I studied power electronics. I studied electrical drives. I studied a lot of like circuitry and so on. None of which, literally none of which, then I went on to use in my further pursuits.

Ashish: Um, even my first job right after IIT was a software development job. and that was also driven by the market economics. Because jobs from electrical engineering itself, [00:09:00] were Quite low paying compared to even the worst paying job from computer science and now that is becoming more and more Apparent to the current students, so they don't even take the chance They just go for computer science even within computer science now people are fine tuning and going for AI Because maybe operating systems is not gonna give you the best edge as a subject

Rui: Yeah. It's interesting how We're a part of the history. Uh, it's just like, we're going through it right now. And then you were just seeing that trend moving and shifting. 

Rui: Um, at what point do you realize that you really don't want to work for the big corporation? Is there any specific that just prompted you to make up your mind of, you know what, this corporate world climbing the ladder isn't really my thing?

Ashish: So my father had a civil engineering job throughout his life in coal India limited. as you will imagine, our lives were always somewhere close to mining area.

Ashish: So we always lived in [00:10:00] remote places. So I'm trying to picture myself, Ashish, who. Just graduated from IIT Roorkee. Full of energy, full of big ideas, and now gets to live for the first time in a big city, Bangalore. 

Ashish: highly energetic, full of ambition, that I'll do this, I'll do that. And the next thing you realize is that, the entire huge machinery is, it's so big. Um, I do remember one particular time period when, you know, code freezes, right?

Ashish: So big companies have these like releases. There's a code freeze. And I think it was almost a week, I was essentially not doing anything for the entire week. After doing nothing for a week, the next Monday you resume, and you just work and then you make some code fixes and so on.

Ashish: but it really doesn't impact the bottom line, nowhere close to impacting anyone, anybody. Yeah, there are many, who would enjoy this, right? Because this maybe makes for a great work life balance, that you could do this, [00:11:00] for your lifetime. And maybe at a different stage of life, this makes more sense.

Ashish: But at least for me, it did not make any sense. I was, bursting with energy, to do more stuff. Uh, maybe that is where I was so frustrated that I would go back home, and I remember, I was finishing. Coursera courses, one course after the other, every other day or every other two days.

Ashish: it's basically the, um, you know, the feeling of, I am not making any difference in this. Like, I'm just a cog in the wheel. it just made me feel I would rather take a huge risk, go to a very small organization where I can at least take on the different responsibilities and then see whether I'm capable of that or not.

Ashish: a mix of all of these things led me to believe, hey, let's try startups. And then I never went back.

Rui: Interesting. Um, it sounds like a lot of the motivation is that you want to feel like you are making an impact and making a difference. 

Rui: I'm just [00:12:00] throwing you a curveball and maybe we're just too quickly tumbling down an existential crisis, but at the end of the day, if you really think about it, whatever we do in this lifetime, a lot of times, You really don't know what 

Rui: you've done that actually made a difference, uh, and in the eyes of the universe, it doesn't really matter. So

Ashish: That is true.

Rui: did you make a conscious choice of not being nihilistic you know what, I can make a difference and I'm just going to keep going at it.

Ashish: No, these are like great questions, right? you're right, like, you got one life, and you do what, works best for you. when I was at Sony, my supervisor asked me, 

Ashish: are you enjoying, your master's thesis here? And I was like, yeah. And he said, okay, that's all that matters. If you are enjoying this, you will do great. Like maybe I enjoy, being, in command, taking control and being responsible, being held accountable, like making a change visibly.

Ashish: Uh, part of, this is also about 

Ashish: when a company becomes [00:13:00] decently large, um, you tend to go in the managerial mode. And everything becomes a bit more inefficient. 

Ashish: I was just in India for the whole month of December and I was helping my parents with a lot of their admin stuff around banks and so on.

Ashish: And I had to go to a public bank. And I can tell you The entrepreneur in me was kind of in agony looking at how inefficient things are. So I had to change a mobile number in one of my mother's bank account. I had to fill a form. I had to change her address for the same bank account. 

Ashish: I was asked to get another copy of the same form and fill another form. Just for no reason, apparently, it's just a thing that goes on, right? So these inefficiencies of large organizations could Disable you from making the same impact that you could make standalone. it's a bit like [00:14:00] if I know I have to go somewhere and I have two modes, one is a bus that will stop everywhere.

Ashish: That will take a lot of people, but I have a bike with me as well, which I know I will have to pedal, work hard to get to that place. I am the kind of person who will hundred out of a hundred, not even 99 out of a hundred times take the cycle. And just ride it myself rather than just sitting on a bus and let's see when the bus takes me to the other destination.

Rui: Yeah, that makes sense. it's about who you are and what you're really in general drawn towards. you mentioned the startup experience in Belgium and also Revolut are two professional experiences that are really, formative for you.

Rui: Before we. dive into those experiences. I also noticed that you worked for a non profit Organization it seems like in Belgium. It's called Thought for Food. Would you like to talk a little bit more about that?

Ashish: yeah, sure. it actually started when I was in IIT Roorkee. a bunch of my friends, we [00:15:00] were actually participating in a competition that was organized by Thought4Food. 

Ashish: One problem was food wastage so We built an app and we actually worked on some ops around to Tie up with some restaurants in Bangalore.

Ashish: We showed some traction, in the sense of. Hey, there are some restaurants which are going to throw away food at the end of every day and through our app they will in fact donate the food to us, notify us and we collect the food and some organizations will then feed it to someone in need We were actually selected as one of the finalists for this competition.

Ashish: So we went to Berlin presented the idea, and so on. And from there on, I chose to become an ambassador with TFF for the next few years while I was in Belgium as well. And we were essentially coaching participants, in thought for food, on their ideas on what they are building, around anything around agri [00:16:00] tech or food tech.

Rui: What was something that's interesting? That you learn along this journey about, you know, food waste or agriculture tech, 

Ashish: So I have a lot of, uh, theories around agri tech because I have been continuing to work on agri tech on the side. it's one of those verticals, of industry in this world where really not enough good people are going to make impact because there is a cold start problem I'll give you the other extreme, which is fintech.

Ashish: Maybe FinTech doesn't need that many machine learning engineers, but it just has the money to attract all of them. And so it gets all of them. And the other way around is true for AgriTech. It has actually a lot of great problems to be solved, which are actually technologically challenging, which will, make you feel so good if you solve them.

Ashish: Very basic problems, but there is no money in it. Um, there are long [00:17:00] cycles. I mean, if you were an investor, you would want to see the promise of a return. And the cycles are way too long. 

Ashish: And so, Agritech is hard to build a business in. It needs a hell, lot of conviction, to pursue. right along the time when I was working at Revolute, I also became an advisor to one of my best friends, uh, startup.

Ashish: And I have been with him in his journey from day zero. The startup is called Swind and they started building software for drones, but then they eventually have pivoted to building drones for Acritech. and the journey is long. It's arduous. my friend who is the CEO has a lot of conviction in this, which is rare to see, which I applaud. and hence he's pulling off the business that he's able to pull off. there are very basic, things to be done, such as spraying pesticide on mass in very difficult terrains, which humans cannot perhaps do at scale, or maybe it's dangerous for humans [00:18:00] to do so.

Ashish: trying to detect any stray animals coming into the crops, destroying the crops, which can again be done using IoT, and AI. all of these are great problems to be solved, makes real impact, but, such is the macroeconomics of the game, that it hasn't really reached the point of, flourishing 

Rui: I love this perspective, because essentially what you're trying to talk about is the structural Problem of this business, right, like the environment and the nature of the things and why it makes it difficult to attract, you know, investors and therefore tech talents and to put in time and effort to really solve those issues that are actually fundamental to people's life because it's food, 

Ashish: Yeah, it's as fundamental as it can get.

Rui: Right. Exactly. So have you thought about like how to nudge specific parts of the structure in a way that you think would actually accelerate, uh, the inflow of investment and potentially talents?

Ashish: Yeah. [00:19:00] Absolutely. So first, the macroeconomic movements. there has to be a external push which is being made by governments across the world, uh, by. giving grants to such startups because they are working more for a larger cause Also VCS.

Ashish: There are a lot of VCS that are coming up that are more around sustainable energy renewable energy Similarly agritech. So there is an ecosystem developing for this with the long term game in mind And so that's more top to bottom Right. The bottom to top here is If you think about the existing workforce within agriculture, compared to existing workforce within finance, finance people, at least we're used to using Excel and they have slowly been ramping themselves up to using Python and then chat GPT comes about. Now they don't even need to use Python. but think about the agricultural workforce, [00:20:00] farmers, 

Ashish: So what needs to happen bottom to top is the education of that entire workforce, which is what my friend is actually doing in large parts to Not only sell something to the existing players, but also to educate them to be able to use the new technology. So a lot of startups in Agritech not only have to build a product, but they also need to build an ops because there is no one who they can deploy the product to.

Ashish: If you see my point, because there's a huge gap. there are a lot of companies, by the way, in Agritech. Which are just ops companies. They are the middlemen between the uneducated recipient of that product and the ultra educated provider of the product. And I think over time that sharp gradient is slowly diminishing having access to learning more about how these things are done. Everyone, for instance, is increasingly having a mobile phone, which is a [00:21:00] basic smartphone, which has a screen, there are, on social media, there are people who are educating. Even my friend, again, has a social media wing where he's creating educational material.

Ashish: So a lot of customer education needs to happen, from bottom to up, and then from top to bottom. A lot of, external push has to be made because implicitly, if you were just to go after money, you would keep pushing for companies like NVIDIA. 

Rui: It's really interesting to think about more traditional industries because because my dad, for example, he is a professor in minerals. And so he is in this traditional Industry engineering side of the things. And I also see there is a gap between, people in that industry who are used to do certain things and they don't really catch up with all the technological improvements that people in, let's say in fintech are much more familiar with, right? So there is a huge [00:22:00] knowledge gap.

Rui: I totally see why your friend not only sell products, but also just create an educational material. 

Rui: I have to say that is a really tough. Business to maintain, 

Ashish: 100%. It is, it needs a lot of conviction. You do not get rewarded frequently. and worst of all, like when you need to fundraise. you get respect, right? But respect doesn't pay you the bills and the salaries of your employee. So, it's a hard problem to solve, but I think you need, hence you need, like, strong people, uh, to come forth and also solve such problems.

Rui: Yeah, for sure. the first episode I recorded for this podcast, Tristan, he worked for a social enterprise, in forestry. In Kenya, the first reaction I had was like, but forestry probably takes longer than agricultural products 

Rui: And that social enterprise failed because of this long cycle. And also the investors don't really agree on exactly what will be the, vision, how the company really makes money and how long does that really take? So I definitely [00:23:00] think investors will have to figure out a mode. To sustain that and that will take time and it will have to get off from those software dopamine where the cycle is much shorter and you get huge unicorn in very short period of time.

Ashish: Yeah, thinking of us as collective beings, or collective intelligence. So You spoke about that in the end what how does it matter? What do we do in this life? And so the other aspect of that is there are people like these who keep trying even if they fail They are actually making dent into the huge Huge, thick surface that needs to be cracked.

Ashish: And so we need more of such dents, and eventually we hit, you know, a crack and then the whole thing gets open, which is, yeah, it's harder to do in some industries than other, uh, but eventually has to happen. 

Rui: for sure. Um, now maybe let's talk about the startup that you worked on, in Belgium. how do you really pronounce that startup's name? Like I see its name, but I don't really know how to pronounce it.

Ashish: So it's [00:24:00] supposed to be pronounced as sentient, I think that was the origin for the name as well, which is to create something of sentience. this was from 2016, if you just picture that in context, this was a time when there was no transformers.

Ashish: At least in 2016, the paper came out in 2017. this startup has been focusing on trying to get user insights. So, imagine they build an SDK that any company like Uber or any consumer facing company could install in their app.

Ashish: And then basically getting a lot of predictive insights about the users. Something that maybe thousands, tens of thousands of companies have since been trying to do, right? to enrich user data. I did a lot in those three and a half years, basically working on a lot of GPS data, which was new to me.

Ashish: trying to understand how to get features out of geospatial data based on the fingerprints of, user, a lot of, uh, Motion sensors data, so accelerometer, [00:25:00] gyroscope, that again is a lot about signal processing and Kalman filtering.

Ashish: one of the things that I feel proud of Doing in my time at Sentience was LSTMs were a huge deal Back in the day long short term memory cells this was the state of the art in Anything sequential or text, uh, data before transformers became so huge.

Ashish: and before chat GPT blew up and I worked on a project where one of the clients was trying to ask us about interpretability. Why are you saying that this user will likely go to the gym or is seeking sports? Why are you saying that? I was able to build like a interpretability layer on top of LSTM, which was visualizing all the different cell activations, based on the different inputs, basically representing where this insight for sport is coming [00:26:00] from, which cell is activating, which cell is the sport cell, basically. in the model, So interpretability is increasingly getting difficult as the models keep becoming bigger in size. And maybe that was my last time when I ever got so deep into the inner workings of the model. 

Rui: Interesting. Did you customize the entire interpretability layer to the model?

Ashish: Yeah, so with LSTMs, Those conventional methods around like SHAP they are hard to, visualize because you have to open up the LSTM cell across all time dimensions. And so the real question is how do you visualize each time the gradient passes through all the cell gates and so on?

Ashish: we actually did a simulation of sorts where I would pass 10 different user timelines through the same model and 10 very contrasting users to show which part of the model is being activated on which kind of user. So, this was more of a [00:27:00] visualization play on the entire model, which is maybe 64, these numbers might sound stupid now, but 64 was the number of, the dimension of each LSTM cell, which in today's world is more like 128k.

Ashish: So, it was still possible to visualize out of the 64, where are you seeing the highest gradients, across time when you pass this user's timeline. So let's say that's index number three and that user does a lot of sports. Now you pass someone Who doesn't go out on a weekend at all, based on their past data, and you suddenly see indexes 9 and 11 are the most, fluctuating, or have the highest gradient.

Ashish: So, basically be able to show which LSTM cells are accounting for which particular behavior, and then also show consistency across users, because this particular client, was very, cautious. About why you are making this decision because we have to [00:28:00] then make some insurance based, prediction 

Ashish: so yeah, that, that was one of the memorable, I would say experience for me to work on.

Rui: So basically you customize that entire integrated ability layer just for that model and then for the specific client use case, that's really cool.

Rui: Um, so you were in Belgium for three years, Before you moved there, have you thought about how am I supposed to, blend in there? Like, what about friend circle? What about, whether there is a good Indian restaurant there, because you might get homesick.

Ashish: something in me, I don't know, I have this, like lone wolf kind of gene in me, which is, I have become ruthless when it comes to completely leaving everything what I have and going to a different place. And that was one of those instances, actually, even before that, going to Switzerland to was already one such instance, because I was maybe one out of five. Indian students in the [00:29:00] entire thousands of student batch of 2015 in EPFL. So, super alienated and before this I was like the going to parties, working 9 to 5, chilling out with friends, having a lot of friends basically in Bangalore, 

Ashish: So left that to go to Switzerland. And similarly, when I found this job, I had no clue what Antwerp looks like. and the first time I landed there I have zero connections in the place. 

Ashish: I go there then I meet people in the startup and So slowly and slowly figured my way like making friends in the company going out with them and then adjacently making more friends outside of the company.

Ashish: It was slow. honestly, it was also painful at times and one should question like, why would you even go through all of this, but there might have been this sense of, hey, I have come so far to study in Switzerland. I should make something out of this and not just, you know, like, hey, dad, hey, mom, I'm back, back [00:30:00] to square one.

Ashish: And somehow I am fine with solitude in that sense. So, when I moved from Belgium to the UK, that was cakewalk for me, even right now, as we speak, I've had to come to the U S my co founder always. It introduces me to everyone saying that this is his first time in the US.

Ashish: So, usually that is supposed to mean that I don't know what I'm doing here. I feel so lost, but I'm actually quite okay with coming to a place where I have literally no one that I know. Uh, so that's something oddly enough, I am pretty comfortable with.

Rui: Yeah, this is fascinating because I left China when I was 18, for undergrad and I actually lived in Switzerland for three months cause I was trying to figure out whether I wanted to do a international NGO cause that's a completely different career path, right?

Ashish: Which part of Switzerland were you in?

Rui: Uh, Geneva.

Rui: also I was living in New York for a summer for research assistantship. that's also when I realized that I'm not cut out for PhD 

Ashish: nice.[00:31:00] 

Rui: and then I lived in San Francisco for a month, just before I landed in Seattle. 

Rui: I don't know, 

Rui: when you go from one place to another, it's always starting from zero.

Ashish: Hmm. Yeah.

Rui: and it's kind of interesting because I always feel like I am in between space, meaning my identity isn't completely this or that and you're always somewhat an outsider observing, which is a very interesting position to be in because you end up realizing a lot of things about yourself and about things or people that are around you. But at the same time, you're not really particularly one of them. So I think that's a very interesting space to be in and for a long time. I was like, wow This is kind of depressing, isn't it? Like I don't have a sense of belonging anywhere, to any group But I guess I just eventually made peace with [00:32:00] that Realizing, you know what?

Rui: I actually drawn towards that state of being 

Rui: the life that you choose to live and the path that I chose to take on. It's very similar. And 

Rui: there's definitely loss and gain in this entire process is just a matter of like what kind of downside that you can really take on.

Ashish: Yeah. 

Ashish: When you are saying all of this, it's as if you are reading out my life. life. It's pretty similar, uh, indeed. and it's very rare to see this. When I go back to India, I see a lot of my friends, actually my sisters who have established lives in a place. Right. So that's the contrast that, and again, the trade offs 

Ashish: It's a wholesome life. At the same time, they absolutely will, be terrified if they have to make switches like I do, on a timely basis. So, yeah, everything comes at a cost, I suppose. in my case, I'd never planned for any of this.

Ashish: I still don't think I plant too many things in life, I, uh, even as an entrepreneur, we just listen to the market signals and then [00:33:00] we keep building on, towards that in life. I also keep listening to the higher signals in some ways, like wherever life is telling me, uh, we should go. I just go on with it.

Rui: Interesting. How do we really listen to that signal? how do you know that's the signal for you to follow? 

Ashish: I mean, there is something, called gut instinct, right? Like this, there is subconscious unconscious gut instinct. there is some inner voice, which tells you, uh, at this point, you have crossed the threshold of doing this. This has to be the thing to do now. I think everyone has that inner voice, in them. 

Rui: Have you noticed whenever your gut feeling tells you something, where does it come from within your physical body? Have you noticed where that voice really exists?

Ashish: So I'll tell you how I interface with that voice and it could be different for everyone. in my case, there is a wake up from the deep sleep and then there is a [00:34:00] proper physical wake up, which is to get up off of the bed.

Ashish: The first wake up, which is when you just got into your senses, maybe went into light sleep. There is that thin, period of time when I feel I am getting the most clear thoughts from my head. Maybe because the sleep lets you sediment everything, then you get up and then there is this soft transition period where Your entire, being tells you if something is off.

Ashish: Um, if everything is fine, maybe you do not hear anything from within, Maybe you hear about some priority, right? that, you should be focusing on. if you have crossed a threshold of you must take a turn, I do keep getting that signal. 

Rui: Fascinating. 

Rui: at one point you moved to UK 

Rui: you apply for Antler Incubator, and there you need a talent visa. Uh, maybe show a little bit more ordeal about your immigration journey, because I think a lot of startup funders, or in general very talented people, are facing [00:35:00] immigration issues 

Ashish: yeah, I'll tell you one thing. I will copy this from Fei Fei Li, who I just was, uh, privileged to listen to, from one arm distance. she came for one of the sessions at A16Z Speedrun. She mentioned this, and I fully resonate with this. immigrants, are already entrepreneurs. They are doing entrepreneurship in their life.

Ashish: not to,pat ourselves on the back, but this is real,

Ashish: when I see a friend of mine who is based in Bangalore, their life is easy because they were born and brought up there and so, they are literally part of that place. They share all of that culture. There is no friction. They'd never have to think about how people will perceive me, can that person understand me.

Ashish: then coming to the more actual practical stuff, When you are being employed by someone, they control your, right to live in that country which is quite, uh, favorable to the employer, in that way.

Ashish: And that is why, I started writing about anything that I would do in this regard.

Ashish: Getting the [00:36:00] talent visa, getting my PR and getting the British passport. I made sure to document all of that so that if anyone else is Going through this, they have hopefully an easier time, to basically cross these hurdles. 

Rui: one part that I want to understand a little bit more is when the tech talent visa in UK, took some time to process and that basically Make you miss out the cohort of that incubator.

Rui: how did that make you feel at that point in time? were you really angry, anxious or to you it's just, you know what, this is the reality. And. Let me just pivot.

Ashish: Yeah, so I'll tell you the kind of person I am. there is this phrase that lives in my head, rent free, which is first case scenario. anything I do, I always imagine already the worst case scenario that can happen. I live through it, I get sad, and I let it happen in my head already, so that it cannot then happen in [00:37:00] real life and still make me sad, and also, I have gotten so used to being disadvantaged because of being an immigrant that these things would not really bother me. So I still felt net positive that finally I am now able to do whatever I want. So maybe I'll take the next cohort of Antler. 

Rui: interesting.

Rui: I do want to get into the Revolut experience the part thatuh, you helped shape their know your customer KYC system and also fraud detection system.

Rui: Maybe tell me a little bit more about that experience 

Ashish: when I joined, we were supposed to be working on a greenfield project. So Revolut is well known for, they want to build everything in house. let's not use vendors for most things, as many things as possible, 

Rui: sorry, sorry to disrupt you right there. Why do we think that they have this principle and do you agree with that or not?

Ashish: The answer is always going to be, it really depends, when you play the vendor [00:38:00] game, uh, it is true that firstly, you got another link in the chain and you are as strong as your weakest link as they say. So if one thing in the entire pipeline performs poorly, that might not be your own, that might be a vendor that does the thing poorly. It's out of your control. So having control on what you do is paramount, right? Uh, you can build, ship, fix things faster. That's obviously the pro. Um, why is it not possible with everyone is because not everyone is getting the same kind of investor interest as Nick Storansky does. Uh, he has been able to pull off a great amount of private market interest in the company early on.

Ashish: So he had the power to hire more and more people to create those products um, if I want to build like five different things with me and my co founder. We might not be able to So I use [00:39:00] openai for example and that's my vendor.

Ashish: Can I build a foundation model? Theoretically, yes. Should I build a foundation model today to serve my clients? Pragmatically, no. Because it will take up most of my time. I'll need to host it on GCP or AWS. So, it's always a trade off. 

Ashish: And this is something I have learned about what's on Nick's mind. Listening to a podcast, he was speaking on 20VC. Uh, people ask him, when is Revolut gonna do IPO? he said Revolut in a way is already a public company because we are operating in the finance space, so we are heavily regulated by the FDA and I think it's on his mind that we must have full control on our processes because shall we make a mistake or shall we get audited, we should be able to fix things, uh, there shouldn't be any funny business going on inside of us.

Ashish: Consider, for example, data leak. If a customer's data is leaked out through a vendor somewhere else. These are the [00:40:00] risks they cannot take. A video game company, for example, can take such a risk. Because at worst, maybe some of their game artifacts get leaked into the internet.

Ashish: But for a bank, that's actually a violation which might Prevent them from getting a bank license. So I think a mix of many of these things could be the answer. But anyways, going back to the Greenfield project, we were supposed to work on replacing Onfido, a UK based startup which does KYC automation.

Ashish: we did like all of the selfie liveness deep fake detection passport data extraction passport fakeness detection Um, all of that, 

Ashish: So if you think about going from UK and Ireland to more than 70 countries, every country comes with their own different ID document. So to be able to validate and build models that can extract information, validate information, deploying them, at scale, to basically serve more than 40K [00:41:00] signups every single day.

Ashish: And more than 100K logins So it was a great experience in many ways. Um, one is just being able to work at a very high performance culture company. being able to really go into the details of machine learning, engineering, the MLOps, the infrasight of things to be able to deploy huge models.

Ashish: All of those experiences were really helpful. KYC, fraud detection, these are all Risk based elements of any business.

Ashish: I remember spending some late nights, uh, on, catching a ring of Israeli fraudsters, And a lot of fraud attacks happen in rings. So they operate from a place with 500 different mobile devices in that room.

Ashish: And they start taking photos with masks, with mannequins, and so on. Yeah, I've seen some, quite some crazy stuff [00:42:00] while investigating such cases. And so once you get an attack, now this will be part of a bigger ring. So you start inspecting a lot of, associated metadata. We also built a graph based model, which was actually utilizing these attributes of like injecting neighborhood of users to each other.

Ashish: if someone let's say second degree connection or first degree connection has done any mulling. Then you are on a higher risk of doing the same based on that assumption. 

Ashish: Revolut has been a very important part of my working life so far.

Rui: I have so many questions. 

Ashish: Yeah.

Rui: How do you really? understand the value of that specific model product, um, because it's about the hypothetical that you just never observe, right?

Rui: you don't observe the invisible, like the bad traffic that you blocked. And what would have been the amount of loss if [00:43:00] it weren't for your model? So how do you really think about all of those metrics and measurement? 

Ashish: Great question. And I think I was just listening to Mark Zuckerberg speaking with Joe Rogan like few days ago. He is facing literally this problem right now on fact checking. These are very similar problems in the sense, ultimately, statistically, it's a precision recall game. How many good users can you frustrate at the expense of finding all the bad users?

Ashish: firstly, there is a metric to be tracked, which is how much did Revolute lose in a month or in a quarter just by any fraudulent activity, which could be money laundering, which could be mulling, or even Risky, transactions, from first time users, 

Ashish: basically, we always had this north star of minimizing that. So that is a clear objective. if it is a million dollars, uh, the CEO is gonna straight up come and talk to the head of the department, right? 

Ashish: If this is the value for [00:44:00] the quarter, and if we have already hit 50 percent of that in a month, we need to redefine controls. the flip side on the controls is, as I mentioned, how wider a net you want to cast. So, a lot of genuine users will also be blocked in making transactions and you will frustrate them.

Ashish: this can lead to user drop off. So there is this trade off between essentially churn versus safety, churn versus a kind of fraud risk. I will be honest with you, quarter by quarter, there was always a. A cat and mouse game between these two. So some quarters we would have low churn and then 

Ashish: we must reduced the controls in a way that suddenly you have more fraud. Which would lead to increasing our controls leading to high churn. So it was always like an adaptive game you always will have fraudsters learning from what you have built to outdo you. And they outdid us every single time. [00:45:00] So even in the book, I always mention, these are great tools, you should keep on building on them, but just know that the fraudsters will always try to outdo you, and it's a real time game, so you have to learn from what they have now figured to create more controls on top.

Ashish: So yeah, these were the dynamics of the whole fin crime department, and the famous phrase was, no news is good news. So if on a given day, we just open our laptop Slack, We are not hearing anything much from the higher leadership, that's great. That means you're doing really fine.

Rui: Yeah, for sure. I know exactly what you were talking about. maybe talk a little bit more about the graph model that you were building. you know, what was the biggest challenge in that process of building that?

Ashish: Yeah, absolutely. It was a GCN model, Graph Convolutional Network model, which essentially, looks pretty similar to a fully connected model, a neural network, but it also utilizes the neighborhood [00:46:00] information. I will say the biggest problem there was, even at the time of me leaving, was In a graph, you have nodes which are connected to each other with edges.

Ashish: And in our Revolu database, we already had, I think 40 million users. So you can understand where this is going to go. Which is, we had this, huge, like massive, adjacency matrix represented as an adjacency list. so we started off with very basic features for each node in that graph.

Ashish: So you are representing the graph. Basically as node A is connected to node B, C, and E, and so on. But while doing so, you are also representing this A, B, C, and D as feature vectors, right? The bigger the feature vectors are, which is the more the features you can write down for a user, the better this graph model will perform because the more it will find relationships.

Ashish: [00:47:00] And there were two bottlenecks. One was, On a regular basis, we needed to sync this entire graph. As soon as new users came in, you have to update the entire graph again. And we had this adjacency problem where the re running of the entire algorithm on the new users led to a lot of issues because the app was going through updates time after time.

Ashish: New users of six months later have much more features to them than the old users, so the backward compatibility. So the model was never the bottleneck. The bottleneck was the data pipeline that would create the input data for this model to be trained. and because this project was started quite late like already at least 30 million users plus, there was this huge, initial, problem of getting all of this data.

Ashish: synced at any given point of time and then load it in to memory, then train a [00:48:00] GCN model. And then guess what? You want this model to be really dynamic, because any new user that signs up, you want to test that user on this, but that user is not part of the representation of the graph. 

Ashish: And by the time I left, we already saw a huge success in another ring of fraudsters in France. As soon as that first user signed up and committed a fraud of some thousands of euros, which we could not have caught because our other models, which were like static, were not ready to see this new kind of, facial attack. As soon as that person gets added to this graph of, uh, existing users, we know exactly where are the next, you know, incumbent fraudsters going to come from. see, they all sign up, right? They all sign up and they stay silent.

Ashish: Then they reveal themselves. As soon as the first one reveals themselves, imagine that the others are already in that subgraph. as soon as the first [00:49:00] node goes red, we know which one, which ones suddenly have gotten a higher risk to review.

Rui: that is really cool to learn. 

Rui: maybe the last question is just about your current startup. 

Rui: explain a little bit about the product that you were trying to build and how did you come up with the idea and why do you believe in it?

Ashish: Yeah, absolutely. we are building native. Uh, that's the name of the company. We're just two, three months old and at the very high level, the fundamental thesis that we have is I think we could help businesses unlock new markets. that's the principle we are going with.

Ashish: And what does that really mean? at least in the UK or the US, I think a lot of people do not realize. There is a lot to be unlocked with going to other markets. I'll give you an example any TV show that is popular here, like American Idol in the U S guess what?

Ashish: Indian Idol is one of the most watched shows in India, which is a localized version of American Idol. there are countless number of video [00:50:00] games that are really released in the U S or the UK, but they pop somewhere completely different. There are manga comics in Japan written for the Japanese audience, and they are so popular in Europe, even in India, actually.

Ashish: So there are like countless examples as a business You're always asking the question, what next? How can we further increase our revenue?

Ashish: I think that's on the mind of most business owners. There are many ways to do that, right? Launching more product lines, increasing the throughput of what you're doing right now. But there is a huge one, which is, can we unlock new markets? And that is the ultimate motto that we came up with. and localization is all about that.

Ashish: that's where the future is going, I feel. the problem though is There's a huge barrier for a mid-sized or a small-sized company to be able to localize and to go to other markets easily. because it's done manually through vendors, there are companies [00:51:00] which hire a lot of people from different cultures and languages to translate and manually convert the product to other languages.

Ashish: So there is a huge opportunity here to create a kind of a layer, a service, an app, an API that anyone could tap into and get access to the entire world as if. This thing was then built for that market to make things more concrete as a business. as I said, we are just two, three months old, we have found a lot of traction, with our current wedge, which is web comics.

Ashish: So amongst the whole host of things that we could cater to, uh, we started looking into the media and entertainment section and apparently web comics are going through their. A revolution of their own, uh, in some ways. So a lot of comic distributors are, uh, coming up in the market and think of them as Netflix, but not for [00:52:00] movies, for comics.

Ashish: And a little fact we learned about comic industry is pre COVID and post COVID. There was a huge, shift in the market, which is just like movies have distribution channels through cinema, theaters, and so on. comics had distributions through comic bookshops. And post COVID that became the bottleneck because there was no shops, during COVID.

Ashish: And so the industry had to figure out a new way to distribute comics. And that is where a reinvigoration of web comics happened. Obviously we have a huge company called Webtoon, So long story short, we have found like a few customers who are.

Ashish: Asking us, hey, we have this English comic, or we have this manga comic, or we have this manhwa comic. Can you seamlessly make this comic into another language? Which suddenly is no longer just about text to text translation, but is also a lot about understanding the graphics, the styling of the text in one, translating that accurately, preserving the [00:53:00] cultural nuances that should go to the other culture.

Ashish: And seamlessly rendering the same styling in another language altogether. So that is something we have been doing. Uh, there are a couple of examples on our website. We are at usenative. com. and having said that, web comics, are a great starting point, but what we are now increasingly looking at is a very similar use case, yet a huge use case, which is marketing creatives.

Ashish: So if you think about all the marketing posters that are created in one language, it's pretty evident that, it's very similar use case that you have visuals, you have text written in different styles, and how about being able to do that in suddenly in hundred different languages, uh, which is a need of the hour today, like English, uh, believe it or not is only understood by 25 percent of people, on the internet, and uh, maybe it's even more skewed outside of the internet and [00:54:00] that other part of the world, uh, should be also brought in into the market.

Ashish: So that's our goal. and, you know, from my story that I empathize with this a lot. Myself lived in Switzerland that speaks four languages. Belgium speaks three languages come from India that speaks 30 languages. So it's pretty much a resonating, theme in my life and yeah, looking to create something useful that businesses can use and hopefully expand, and get to more markets.

Rui: first of all, I love that comic's story. would you say that most of your customers that you're trying to target are companies that are operating in, like, not even just one or two international market, but like dozens of international [00:55:00] [00:56:00] markets.

Ashish: So there are always in different verticals of industries. There are the big players, the mid tier players, the small tier players. Uh, the big players usually have enough resources that they have dedicated teams to do all of this, like a Coca Cola campaign in English and in Spanish.

Ashish: [00:57:00] I find it hard to believe that they will immediately agree to use us. Because they really care a lot about the final top notch quality. They can spend a few tens of thousands to get the best copywriter to get the best graphics designer. So that's the top tier. Then there is a long tail of like several hundreds of thousands of companies in the lowest tier who, if we just tell them that there is an API, they will just, yeah, let me use this because I have nothing to lose, but they might not be willing to pay too much for it because. They are still finding traction in their own market or they're too small, do not have the budgets to go on into other markets yet. And then there is a mid tier, which is the most lucrative to us. companies that are not like the massive corporations and that are not like indie game developers somewhere in the middle where they have enough resources to, pay for localizing their product or marketing at least to other markets.

Ashish: and yet [00:58:00] they haven't been doing it so far because it's quite painful and tedious. And more expensive than we would offer a service to be to go into other markets. So that is our analysis so far which has been coming true for video game industry also for comic industries And increasingly for the marketing creative ads industry

Rui: finding what's the right customer segment in a way that your product eventually has a sustainable business model is something fascinating, to think about 

Rui: thank you so much for being an open book and share so many different experiences and thoughts, 

Ashish: thank you, Rui. this was really, really, really enjoyable. you have touched upon some points which I have not looked at for a long, long time in my life. 

Rui: Something I'm gonna really, think about [00:59:00] This could be the last episode of floating questions, or it may not be either way. I hope you enjoyed flowing along with us today. If you liked our journey, please consider subscribing. Thank you for listening and made the questions always be with you.