Ep3 - Designing Data Experiments to enhance your Product | Rapido's Data Science Lead, Pramod N

Key Takeaways

i could probably put my head up and say and we're one of the few companies which uh use data at all facets of running a business years as a data scientist you drill so which means you help us understand what is the best way of solving a particular problem at hand [Music] hello everyone welcome back to yet another episode of the data show and this time around i have a special guest from the team rapido now rapido is an indian online bike taxi aggregator based out of bangalore now the company was founded in 2015 and since then they are now operating in over 100 cities across the country now rapido has been incorporating data science across every aspect of their product be it reducing cost optimizing their dispatch algorithm or enhancing the product interface in general now they are running over 20 experiments 20 data experiments in a week and how do they pull this off to discuss this i have promoted with me today who is heading the data science team at rapido now our conversation covers a wide range of topics we have discussed uh from their uh from responsibilities of data scientists at rapido to their hiring philosophy to uh their design uh basically their experiment design philosophy and how they have been able to put together such a passionate team of data scientists so here's the conversation that we had and i hope you guys enjoy it so hello promote uh again i know you're a busy man and thanks for taking out the time to meet today and i'm really excited uh to have you there and talk about the rise of uh applied inaudible is advancing in this direction and how you guys are you know applying data science incorporating those statistical experiments to enhance your product in general so to start off uh like i would like to hear your take on data science as it is such a broad field in itself and there are so many notions around it so like what does data science mean at i rapido it's a very again broad question uh so fundamentally for us uh the vision for us is to be data-driven at all uh avenues possible from a business point of view all right and i could probably put my head up and say and we're one of the few companies which uh use data at all facets of running a business we're a very metrics driven sort of a company and so on it's a metrics driven culture and you know in that kind of an organizational setup data science essentially means uh being able to uh c first game uh like why why we look at metrics as in when we get it being able to foresee and being able to see what what might happen in the future is something that data science would contribute to uh it is also being able to see we are not leaving uh anything unoptimized right like uh leaving money on the table uh not wasting where it is not supposed to waste uh being able to utilize our captain force uh in the best possible way being able to maximize earnings for them being able to give better rights for our customers at a better price and so on and so forth right so data science essentially is there is a there is an organizational setup and a culture of looking at metrics and there are people who look at this at a daily hourly basis uh data science is almost like giving uh rocket fuel to that process right and it's simply there is no other fancy anything going on right yeah how do we people these people to become better and make the business more efficient and provide better experience now be it machine learning or simply just cool statistical experiments it doesn't matter right uh yeah so i think one of the uh i think this is a i would call it a philosophical stands that we have taken essentially right where we are not ai hungry yeah right why we will silently use ai right so we don't want to be harping on the biggest and the fattest algorithm but we would want to be using the right tool for the right job right so it doesn't matter whether there's a simple probability calculation to uh you know like a deep deep learning model running somewhere or there there are ensemble uh of you know different learning algorithms working together i think for us what matters is we are slightly too diagnostic slightly algorithm agnostic again the biggest gun does not give you the better result i think almost like snipers right so even the one more targeted solving the problem is what is more core and more focus than uh you know like using the biggest to the best and that's well put uh the sniper analogy right and and uh so what are the day-to-day responsibilities of a data analyst or a data scientist who is working at rapido right so i think uh there is a bit of background on uh team structure which changes narratives slightly not a lot and i talk about this constantly everywhere like uh so we have a uh we have three legs in a product of wings we talk about the wings the three likes essentially are uh it's almost like uh it's one entity but they have to be played by three people because of the skill sets and what they bring to the table etc uh and the metaphor that i would use to describe this is uh that of a tunnel boring machine i don't know if you've heard of her tomorrow machine uh things that are drilling metros yeah uh other kinds of projects around the world right so there are fundamentally three aspects of this tbm of the machine there is a drill bit which is the front which leads it's like a massive turning thing that you know chugs out mountains and whatever granite in bangalore or whatever and then there is the navigation where when you are underground and you're blind you need to be still able to see through rock whether there is a cavity whether there is hard rock ahead whether there's water so our uh data science team essentially works like a dbm right so there's a product and a goal for the product uh the product almost acts like the navigational piece of the entire uh thing right where as a product owner uh or a product manager you are sort of constantly saying okay drill here drill there there's more value there there's a cavity there you know that's a catch you should not go there or whatever and as a data scientist you are a drill bit right where uh you are facing the hardest of the problems right and you are facing uh the worst of the brand right so you will face the hardest of the rock or the softest of the cavities uh but the expectation from you is you help us push through and drill no matter what right uh and we'll talk a bit about it what does that mean uh where they ensure that once it is just so that you know so this i'll cut out so you don't have to worry about this okay cool so i was just saying uh uh the third aspect of that uh three-legged machine is the uh is the you know arrangement and making sure that after we have drilled yes structure behind this doesn't fall right so our analysts sort of act like a reinforcement to the entire system right where they make sure we are able to run repeatable experiments their fundamental part of every experiment cycle tracking all kinds of metrics uh ensuring they're able to narrate a story of what is happening in an experiment right so so those are the three three uh likes of the journey i think your question was what does it mean to be a data analyst and our data scientist the answer simply is as a data scientist you drill so which means you help us understand uh what is the best way of solving uh a particular problem at hand whether it is pricing optimization around that or dispatch and how do the right people how do you provide a better sort of customer experience because of some of the campaigns that we run or whatever be the case right uh so that that would be as a data scientist and as an analyst your fundamental responsibility is to ensure uh we are able to run experiments track experiments we are able to find problems with experiments and solve for them uh we're able to sort of see where we are going wrong and where we're going right double down on places where we're going right you know stop or pause uh on places where we're going wrong more importantly you're able to tell a story of how our experiment is going right so for example [Music] [Music] one of the like if you just pick that example as a data scientist your responsibility is to find the parts of the problem which with the help of the product and uh you know try and address the part of the problem like for example peak hours we're getting too much demand and we have supplies and then solve for for that through experiments right so whether it is trying out different algorithms trying out different hypothesis uh within an algorithm or whatever be the case now as an analyst uh the responsibility is to sort of ensure every experiment that we pick has a certain business prior we know what was going on before we were able to track are we able to tell a story of how the marketplace was behaving back then right and then there is the conception of the experiment you will work with the data scientist to figure out how do we conceive this is it going to be a b is going to be one team on bandwidth or is going to be abc or whatever be the case right and then post the conception you will help set up and track house experiment performing should we cut it now should be running for longer is it conclusive should we run any kind of uh statistical test to verify whatever is required to sort of conclude so that's broadly how i would say the three roles work more importantly the two roles for data scientist let's all work together okay okay great uh you have a quite a framework there that i have come across for the first time and and uh you know uh the whole drilling experiment that you talked about this is interesting uh how you explain it uh so i was uh like uh uh so i was reading your recent blog on how you guys uh you know uh are designing data experiments to optimize the right dispatch algorithm and reducing cost and and the product in general so if you if you could first give us some context on how uh the rapido write dispatch algorithm how it works and and then how did you go about designing this experiment right so just to give you a bit of context like some of this is quite proprietary but some of this is quite open right okay okay uh trying to avoid what is experimental and what is sort of uh the secret sauce if i were to call it so but uh most of the stuff i can talk about now uh typically look at any on-demand rights kind of a market whether it is a delivery or otherwise uh one of the things that happens is you request a ride on your phone right and you will be assigned a [Music] entire process is what is the boundary of dispatch right okay now what used to happen earlier uh prior to like us uh intervening from a data point etcetera was uh this is a four and a half four five years old system almost uh and it used to do a very rudimentary form of dispatch where uh for every request that i you know create uh it's almost like there is an imaginary radius around me let's say one two or many kilometers right and whichever captain is in that radius you will start picking them you know are you interested in this order are you just disorder roughly that's what is so there are certain issues with such a uh like a methodology of dispatch right so one is yes it happens the other is the radius every time you draw this radius it's very unique for a person simply because it is radius around the laptop right so what that means is just from a sourcing of who do you take a ride with that becomes slightly less predictable right so if you're just a few meters away from this uh arbitrary dispatch radius you probably are not part of the you know like the dispatch queue recorded yeah so and these things make the marketplace uh slightly more inefficient right so because everyone gets a new radius you know sometimes the radius has uh lakes as half the uh you know half of the circle is late right how do you do that right so then from the marketplace point of view it's very important so our first step in changing this was to ensure we don't just draw arbitrary radius we make the uh entire marketplace slightly more discrete and more predictable right so what do i mean by that so currently uh if you see i don't know if you've worked with the uber's h3 library and what uber did to solve very similar problems and they've open sources open source this library and it's a fabulous piece of engineering just from a thinking and how it is architected uh so what as a first step one of the first things that we have done is to adopt uh very similar thinking as to how we will look at jio how we divide jio so currently by doing this you know point and radius and so on uh we almost made geography as the continuous space right yeah like every point is new it is uh you know has less order it has uh you know no way of folding and no way of looping and so on and so forth so which made uh us gain any efficiency sort of slightly harder right so so the first step was to make that more discreet so what do i mean by that uh we use the h3 concept or the hexis concept to sort of discretize our jio where now it's not a radius around every person but it's uh it's almost like a radius fixed geography right so okay as a consequence because of this is you can eliminate things that are uh you know like lakes and uh that are military areas and so on and so forth uh which you will never even consider uh to be part of uh sourcing or dispatch right so this is a very small example of uh like the first most like smallest can you hear me now this is any button yeah it's fine okay sure uh yeah so this is just this is an example of a very small change uh which brought about a lot more efficiency in the marketplace than uh before right so what this has allowed us to do is to uh like do a lot more experiments on how we source how we group jio and how we manage supply demand levels purely by managing the adjacencies of uh these hex right so the experiments have primarily been how to uh go about grouping how to go about sourcing when do we do what kind of sourcing when do we do what kind of grouping uh should we uh you know be predicting at a hex level should be predicting at some other level you know some of those decisions is why we're doing a lot of these experiments the primary goal of these experiments is to reduce ets right so from a customer point uh what we don't want our customers to do is wait internally right uh in the pursuit of reducing eds we are trying out a lot of these experiments and the second part of this is to sort of reduce uh a burden on our captains where uh if there is a dispatch and let's say people are far away they have to travel longer to pick up someone right uh yeah right so the ultimate goal of this dispatch has been to sort of solve for these from these higher order goals which are reduces reduce eta uh improve uh earnings from a captain point of view uh and that as a consequence like these things should happen as a consequence of better matching right so that's what dispatch experiments are all about and i think in the blog you saw just one example where yeah if we replace some hex with another because of yeah yeah and so on like you will gain some efficiency i think that's just one example yeah uh so also what were the challenges uh while you were designing these experiments or running these experiments uh one of the primary challenges uh it's hard to do test and control like it's hard to do av or anything yeah simply because uh because these geographies are not homogeneous and it's hard to define modernity of geography uh what is similar what is not uh defining test group and control group are which are not at the people level which are fairly discrete yeah uh is another fundamental challenge right so how do you make sure that something that is uh less controllable uh become more controllable i think that's the biggest challenge of uh setting up these experiments i think in the blog we talked about one such example of how do we how we came up with some of the parameters which define geography where we were able to see some geographies as being similar from a need point of view uh versus others uh and that allowed us to sort of separate cleanly slightly more cleanly who's taking ride why are they taking ride uh more importantly why are they taking right then uh who's taking right okay right so i think that's the biggest challenge with running our dispatches so does it also include like what time of day it is and and based on that you decide like uh uh who will be exposed to what sort of experiment or what sort of trial yeah i think the most of the problems in the you know marketplace kind of a setup is always your temporal in our case right when i say geotemporal it is a combination of uh geography location as well as combination of time so morning peaks are not the same as evening peaks yeah afternoon people different night is different and so on and so forth right so very similarly the same location let's say it's a very hot command center in the morning but might not be a demand center in the evening right so it's definitely a geotemporal uh like kind of an experiment set up for sure okay okay so like the azure we are undergoing these crisis and i want to know like what did you guys do when the entire world was under lockdown and and rapido must have not had any business back then with no rights and how did your goals or your vision change at that point of time if at all they changed uh so i think just from a vision point of view like as a company we are here to move and solve movement problem for india right yeah uh so bike taxi is one of such examples i think now we've started doing motto uh we've been in the delivery support sort of uh like logistics board sort of a business anyway so i think movement anything that we move i think we've also launched a couple of other products but anything which involves movement we wanted to be able to uh sort of contribute right so because india has a lot of opportunities purely from like transportation movement of people or groups uh point of view uh our current focus has been to do all of this hyperlocal right and we will continue what we have done is we launched three new products uh our offerings uh one of them was a courier offering where yeah during lockdown if people wanted to send stuff to each other you know they could do that which was again aligned to the vision of you know one more people or things in india again we launched another product on uh on auto which was again aligned to the very same with similar supervision we launched another product on uh groceries but not like everyone else is doing i think ours is slightly more nimble and more niche i would say uh because again even a fundamental evolution like now we need to build it in bits not big bang products or whatever uh so which means like the messages uh in the past six months while we have been worried uh about like the situation and what is going on etc we've always also been very more focused okay what do we really want to achieve how do we want to get there should we just completely take a u-turn should we do something else like what is that and i think the answer is emerge at least as a leadership team uh uh like all of us go through this what we're seeing unanimously across the board is that we want to head more closer to our vision uh and build uh so that when you know things are slightly more normal we'll be able to execute better as well as uh solve for problems right now like i think a lot of people asked for a slightly more safer way of doing travel so we launched an initiative on the shield i don't think anybody else in india is doing that uh and so on and so forth right so i think we while we focus on the current need we always have the view of the future of what where we want to be what we want to achieve and i think that's what has happened in the past six months right so this is purely from a business point of view yeah if you look at other things i think people one of the uh conservatives will we will not uh randomly fire we built a very interesting uh team yeah it has taken us long time to build a team right so we've unnecessarily not let go of anyone has been to make sure that we double down on the team we get more out of the team uh putting together intelligent people and then saying you know what now it's not required it's probably uh not as uh ideal as we'd expect it to be and hence we've said okay we'll not uh get rid of but instead we will double down and we'll do more right so i think that is another change that has happened over the past six months uh and finally i think what it is giving us is a conviction that if we ride this wave while it is a lower slope and it's a harder situation to be and so on uh it has given us more conviction that this lot more is possible and we were going at a certain rate and then we had to suddenly like there was a cliff and we had to stop now we are i think we are almost in the middle of the clip i don't know if you were done with this yet uh and now we have the computer okay we can we can face this uh even stronger and we're hoping for our next quarter to be slightly more better so uh we worked with a lot of government agencies uh in the lockdown and did some social good there and now we're looking forward to sort of doubling down and expanding rapidly what we were doing okay okay that's interesting uh about uh you know the letting go well there were people being fired so as you know there are five million jobs that are gone during and in the organized sector and the remaining the number is just a very low very small number for what the ground reality is so yeah uh great uh and so is there a data science philosophy uh the team philosophy that you guys follow within the team like this is what the data science team is for or like the mission statement or something like that that drives the team uh mission i think mission statements are for a mission like i think our mission is endless okay uh the way we put it is the the biggest data science philosophy for us is experiment experiment expert like we want to call that commission statement so right so which means we will never sit in a corner build a model and be happy about right i think for us in fact like some of the measures on some of the products is the number of experiments we do right like how many experiments are we running and how are we able to conclude experiments whether it is success or failure uh i think that is a measure for like it's almost there for some people right so i think that's i would say that is the most underlying philosophy of what we're trying to achieve which is learn through experiments do it faster fail faster uh don't you know model for the sake of modeling experiment driven is probably one of the ways of putting that yeah okay interesting now i i get that you uh you have or have been working on a lot of different kinds of problems and uh using data to solve them but like uh i'm sure not all of them would have been successful in the first go uh so can you can you walk us through like one of the failures that your team didn't see coming like except the government thing that no one said so right so i think failures we have failures on a daily basis right okay and i think the part of experiment is to embrace failure than avoid failures right so an experiment essentially means you don't know what is going to happen yeah i'm okay with either success or failure i will get the learning what works what doesn't work all right so which means if we've taken an experiment first approach it's not like all experiments always succeed right so like a lot of experiments pay and that is part of the job and we don't even look at it as like a setback because at least we now know that it doesn't work right so there's no point fretting about uh not making it work you know yeah we pick up pick up another one the reason why we're able to do that is uh twofold one is uh we ensure that the running of experiment is cheaper like there is a constant focus on if we constantly run cheaper experiments the more costlier the experiment uh there is the failures become costly right the only way of sort of getting around failure is to reduce the cost of failure right so if the failure cost is very low or how does it run right so i think that's given that kind of a approach in the philosophy uh i think we fail very frequently and more frequently than a lot of people can imagine like if you run an experiment a week i'm pretty sure there is a failure every week right because we are able to control the size of failure and the cost of failure yeah it does not impact as much and is not like a massive setback right so just one of the examples was we ran back in feb we had a certain hypothesis of how to identify segments or and stuff like that i think at some point we were running 20 experiments a week right just on that product and i think one of our data scientists said actually this is not working because you know we're not defending statistical significance and we had to discard like a few weeks of experiments uh as a consequence but we already had an alternative because we knew knew what was not working right and these things constantly happen right because we had an alternative we knew what was not working the experiments were unclean et cetera tweaking and changing became very easy in fact that led to the one of the optimization solutions that we built on how to invest in our customers and capitals right i don't think [Music] reduce the cost of your experiment make your experiments cheaper uh fail more so that you're able to fail cheap rather than you know waiting and waiting internally finessing the model or whatever uh that philosophy has helped us sort of avoid a lot of setbacks even the setback that i was talking about is just uh it's a matter of uh a couple of weeks of effort and we were able to pick it up and come again and we were able to run faster than uh what it took us to get right and does it include the time window as well the experiment like should it be like uh you know let's say i would run this experiment for just four weeks or six weeks uh so it should be like uh the time window should be shorter longer or is there any take on that like i know that a statistical experiment should be should be kept uh like uh for a short window of time but uh again uh i don't want to put a number on there because that's that's not correct and the reality is going advanced and and it depends a lot on the problem statement and what you are about to your hypothesis basically yeah i think you're right like uh uh there is a general guideline of smell of learning about experiment okay uh but we don't prescribe what would be the direction so i think there are a lot of online letters as to how uh how much longer like what should be the size of the sample and the duration of the experiment yeah which sort of give you significance eligibility so i think we use we use some of those like we use some of the methods to sort of uh make sure that we're getting enough samples and so on but again in the spirit of running cheaper experiments and uh doing faster cycles i think the way we set up our experiments has a lot of bearing in this right and sometimes it's a constant debate look this costs us more but uh the return is not so much should we even try this like should we do some something suboptimal uh and so on and so forth like this is a constant debate and one of the challenges also should we run longer experiments right uh so i think currently where we are at we've left it at uh there is a general guideline of if your experiment is running for more than two weeks yeah it's a smell that's something that something is not right uh but we have run more than two weeks uh experiments uh in certain scenarios right so it's not like we stop it but it should uh be an alarm they're saying okay something is wrong or should i look at something more and so on right so i don't think we have a fixed timeline whatever i think we follow this general guideline uh an experiment should not go overboard okay okay fine i think uh uh so my next question would be like what do you look for in a candidate you talked about you know you put together a really good team of like really intelligent people so what do you look for in a candidate who is applied for this position of data analysts or data scientists uh so let me probably talk about what we don't look for right like we're [Music] yeah we really appreciate the fact that you have the uh transformations and through this like great yeah you'll love it uh but i think what we look for like a few of these uh things right so firstly you've got to be curious yeah as a data scientist you're you're the drill bit right yeah uh being a little bit means taking a lot of heat uh and if you're not curious you will not be self driven and you can't take that heat otherwise right yeah so if you're not curious about what is going on why is it happening happening and so on and so forth and if you want constantly for someone to be telling you i should be doing this versus that yeah right so curiosity replaces nothing i mean you you might know 100 algorithms sorry curiosity is not about i want to understand or learn an algorithm curiosity is more about how does this problem work can i think about the problem in a different way right can i appreciate the constraints of a problem can i appreciate constraints of the solution and so on and so forth right so this is not about curiosity doesn't mean like i know 20 000 meters yeah okay so i mean we are at an information age where click of a button if you know the higher order concept you will get the details of the methods right while it might be slightly inefficient we are happy to take that learning and onboard that inefficiency rather than lack of curiosity on problem thinking about the problems breaking down the problem you know thinking about constraints for the so the biggest thing that we look for is that curiosity of how do you think about the problem uh do you think about it holistically the second thing that we look for is how is your fundamentals so as a data scientist one of the expectations it is that we have with you are the master explainer like you you're able to tell a story uh which is not just uh understood by your fellow data scientists or analysts or you know product people but it is also understood by people who don't have anything like this like whether it's people in ops marketing whatever else so the idea behind a data scientist being able to explain and tell a story is to also make some of the analysis more palatable for anyone who is sort of coming into this process right i think what we have observed at least in my past experiences explainability improves plus uh whenever you're building operational levels etcetera uh whenever it becomes a black box the trust in gaining trust takes time takes more time right so you need a certain level of expertise to appreciate the black black boxness rather than the explainability of why given most logical things can be explained right so that's the second like you should be able to you should have a core fundamental uh which allows you to explain why is happening right yeah the third thing that we look for is uh a lot of patients and being able to skim through a lot of data if you're not interested in looking at data and you're only interested in modeling when i say modeling like you're only interested in creating model yeah i don't think right so i think you should genuinely be interested in looking at correlations looking at all kinds of slice and dice looking at things like will aggregation change we'll have to look at a different pdf and so on and so forth like just being able to spend a lot of time and have patience to spend a lot of time uh with that i think is another fundamental quality uh characteristic that we look at the reason why i call out these things uh things is if you use something like an h2o or any other online modeling like the uh dynamic and ml engineer modding methods like the tools etc they can do half the job which i just described right the reason why we need data scientists or data analysts because okay that job is done what does that mean like how is it applicable what is working what is not working i think that is the job yeah right so job is no more how do i build a model i think building models nowadays is a bit uh it's a bit like saying i build a compiler right it's fine you know the things have standardized there are some you know tweaks here and there etc but i think at least in our stage where we are right now uh learnings and being able to explain and being curious and looking at a problem differently et cetera that is what we value more uh rather than just having an understanding right so so if you ask me if you're applying for rapido this is what you would have gone through okay yeah hopefully that would be uh of value to many more candidates who are looking at this yeah and and one other question that i have is again very controversial so how much how much does a degree play a role in shortly in the shortlisting process uh so i think one of the things that we encouraged some of them are ex-data scientists some of them are like people who've run a business there's a tech lead in there this is where product group very similarly in the data science group uh there are very core statistical background people there is uh there was a developer and there's a quantz guy and so on and so forth right so i think from an education and an experienced point of view uh we are less uptight about it if i have to call it something right so you're an engineer and then you have an analytical uh mind and uh you have a strong fundamental and you're okay i don't have an issue very similarly we've done some other course or degree or whatever that should still be okay i think what we prefer is some proof of having a strong statistical fundamentals right so while our assignment acts like a proxy uh there is only so much you can assess through one problem right so yeah okay look i have a uh statistical engine i have a vertical mind and i know the fundamentals of uh how to solve problem using data right so i think that's that's what we focus on so education uh it tends to act like a proxy like if you have done uh you know msn computer science and statistics or whatever like it tends to act like that black proxy or look i i know some things i don't just lying or whatever i think that's the only purpose education sort of like the qualification sort of uh gives us uh but it's definitely not uh shall we say uh replacement for having analytical uh so i think we have had instances where we have rejected ms candidates and we have we have data scientists who have nowhere ever been in any kind of a statistical heavy course right so we have engineers who took this up and down like i think we have both examples so the way we look at education is it's an indicator it's a proxy for something uh we just use it as that uh it is not a hard and stop uh rule or a boundary okay okay that's good and and i want to know like before rapido you were also you know you were working for a service-based uh company uh so is there is there a difference in the incorporation of data science uh in a service-based company as compared to a product-based company yeah i think one of the reasons why i wanted to sort of come to rapido was also to prove that point i suppose i think service-based companies there's nothing wrong i think they have very similar opportunities in fact there are more opportunities there looking at different parts because different kinds of businesses different levels of intensive problems and whatnot i think the only uh difference i see is uh as a leader there a lot of the time went in uh convincing people that they should look at metrics uh they should consider these uh you know experiment driven approaches and so on and so forth given the nature of organization and the nature of engagements and so on a lot of the energy was just spent in that rather than doing it right so the only difference i see at least personally the only difference i see is uh there i was convincing about doing something here i'm actually doing something and then figuring out you know what are the applications i think that's the only difference otherwise i think service industry currently given the demand for solving a lot of data centric problems there's immense opportunities across the world uh it doesn't matter technically it doesn't matter where you are i think the only advantage you will have in a product-based setup is if you have a right leader and a right philosophy and a right team you might be able to do a lot of these experiments and things and a lot of uh lot more data-driven productions i think i think one probable reason could be uh that you you said that you know uh that company was maybe very tightly organized okay so that's why you would you had to put in a lot of work to convince people to do something and and you'll be you're at a startup now where you have a lot more flexibility uh to take up and think more in an entrepreneurial fashion so you know that probably could be one of the reason why you actually don't have to work so much to convince people to do something or look at a few set of metrics uh and then yeah again uh great so uh my next round is the uh rapid fire on where i'll just uh throw some really quick questions and you just have to answer like within like let's say five seconds so okay okay let's go uh first question r or python or sequel uh i think this is controversial so i would say python uh given my background okay okay one thing you love about working at rapido the hustle like the constant numbers and hustle okay and one thing you would like to change about rapido uh i think we could uh uh become slightly more uh empathetic towards uh a lot of things like i mean this is not a again this is a constructive criticism where yeah yeah very extremely matrix right like yes like almost all conversations are numbers right okay uh even a lot of empathetic things happen through numbers which is just good and bad so i think we can be a lot more empathetic about uh you know how customers feed how captains feel and there is a lot of that already right like uh one of the fundamental uh things about both like what founders have done and like what the current leadership is doing et cetera like all of us constantly talk about hey do we have a voice of the customer we know what the captain wants and so on and so forth uh but i think we can we can do a lot better i think the bar is at least from my point of view the bar is slightly higher and i think we can get that okay the go to learning resource uh that you advise every incoming junior at rapido to learn from sorry say that again what was the first bit go to learning resources that you advise any incoming data scientist or analyst uh right so i think uh currently the due tech and like a lot of the online information systems like the way they've grown there's so many uh uh but i would always definitely advise anyone who's trying to do something in mla like uh android is a great place to start okay uh was also one of the master explainers like he was able to explain a lot of things uh etc uh he was one of the first people who sort of probably started the ai ml and started yeah uh doing online courses back in 2012 uh and uh there are also a few classic books uh peter norway uh sebastian run uh there's a there's a three author book on artificial intelligence like this part yeah talk about a lot of uh cool things and talks about a lot about fundamental stuff like i think currently if you look at the ai narrative it is almost always about machine learning okay i would recommend aspirants even students start with that book it's uh it's slightly dry uh but it gives it opens uh you're like thinking up for a lot more ways of thinking about uh how to do fuzzy right how to do yeah intelligence in general right so uh again it's a dry book uh caveats i did but definitely a good place to start uh apart from those i think there are so many courses university of uh you know john hoping hopkins has a lot of courses you know florida has a lot of courses stanford and others have also mit has started like there's so many things going on uh i think there's no dirt for learning whatever i think if you ask me personally what are the two things that i would do yeah uh i think supporting these two uh for sure uh in addition to i think there's a good online community on data science central and uh getting nuggets yeah if you guys have yeah that's a i think they are also a fantastic resource for to just learn from others like while this is learning theoretical and fundamentals just learning from others those might be so uh learning or core statistical experiment design uh which would be preferred to master is there a third option of don't care but master role i think like the way we are yeah like the way i look at it is uh uh i think it's the right tool for the right job as long as you know the tool okay so if you were to start learning data science all over again uh which part of the learning process would be changed other way around i think that's why i have an appreciation of why you should do fundamentals first so i started in a lab which had nothing to do with the statistics or whatever we started applying machine learning even without knowing a lot of the basic learning i had to learn basics as a consequence of not understanding what is going on right quite honestly so because like the way i think and the way i uh solve and the way i learn is more ground up and fundamentals first i had to go pick that up and then i realized oh there's so much to go and do uh if i were to change that uh maybe i would do the other way around okay okay so that's it uh one last question uh before we end this call so do you have any parting message for an aspiring data scientist or any data professional in them like what should they focus on any anything that you would provide any message uh i think i would repeat what i said earlier like there is no replay for being curious right now you stop being curious the day you will become less uh successful data science or any data uh yes entering profession so there is no replacement for being curious and being curious does not mean again methods being curious about problems themselves uh constraints are the problems solutions and strengths of the solutions and so on right there's absolutely no replacement for that if you ask me how do i become curious i'm not a good teacher when it comes to these things but being curious definitely definitely helps there's a massive correlation between some of the successful data centers to how curious they are about problems and solutions yeah great uh well thank you so much for both uh for connecting today and uh i hope uh and i'm sure like this would be uh an enriching conversation for people who would be watching this so thank you so much no problem i it was a pleasure being here i'm sorry about some of the network related things but okay thanks for that all right