Life Insurance Engineering with Vipul Sharma and Lingke Wang


Ethos Life Insurance is a software company that sells life. Insurance Products Software is reshaping established industries such as banking insurance assurance and manufacturing and in these large established industries incumbents are adopting new technology as fast as they can but the new technology needs to be integrated with the old technology and the old business processes. The slow rate of technology adoption by incumbents creates an opportunity opportunity for new companies to spring up who are building entire new companies from scratch with updated software stacks. Insurance is a gigantic Ganic market and it's dominated by companies which have been around for fifty to one hundred years. The established players in the insurance industry are trusted brands but many of them have not significantly updated technology from legacy techniques of pricing risk and risk pricing is really complicated. There's all all kinds of variables the need to go into pricing risk you could build complicated machine learning models and I think it's safe to say that many of these incumbents are not well equipped to make that technological shift. VIP Sharma is the VP of engineering and Linke Wang is the CO founder of Ethos Life Life Insurance and they joined the show to describe the insurance business the technical problems and the software stack of a modern insurance company either has more than fifty employees and it's growing rapidly so it's a great case study in scaling a modern company in an established market email has been around for longer than I've been alive but there's been surprisingly surprisingly little innovation in Inbox Management Sane box is a new way of looking at your inbox that puts features like snoozing and one click unsubscribe and follow up reminders as first class citizens if you are overwhelmed by your inbox and you're almost ready to declare email bankruptcy try out sane box in the on boarding process sandbox analyzes your emails and helps you sort them into categories. You can get a free three fourteen day trial and a twenty five dollar credit by going to sane box dot com slash. Save D that's S. A. N. E. B. O. X. Dot Dot com slash. Sad these days. I spend more time in my inbox than I do in front of my coding environment and back. When I was programming a lot I would would spend hours configuring coating environment because I wanted to maximize productivity if you spend as much time managing email as I do. It's it's crazy not to set yourself up for success with your inbox so stop the craziness get sane with saint box. Good Assane Box Dot Com Sush S. E. D. and get a free fourteen day trial as well as a twenty five dollar credit. Thank you too sane box for being sponsor of Software Engineering Daily Doc Linke Wang and VIP Sharma welcome to Software Engineering Janine daily. Thanks so much thanks in college. I had a strange experience and that was the experience of somebody taking me to dinner and and pitching me on life insurance the whole time this was somebody I had known in high school and he emailed me ostensibly just to catch Gotcha and we had dinner and then he spent the whole dinner pitching me on life insurance. Why did that happen. That's serve a really great question. I had a very similar experience coming out of Undergrad. I met a life insurance agent who was an alum of my school and we met at a bar and he actually convinced me to buy a whole life insurance. When I was twenty two years sold it's a career and these folks are sales people and at the end of the day that's what their job is to try and find people who are could candidates for life insurance and convince them to buy a policy so how much money the sales people make doing that. The average statistic that I've seen online is that a typical agent will make about seventy to ninety percent of the first year premium premium income missions that number is going to vary depending on how many years you've been in the industry what kind of product you're selling so on so forth who buys life insurance the typical person that should be buying life insurance in my opinion are folks who have families folks who have something to protect because at the end of the day the purpose of life insurance is to protect your loved ones financially. The worst thing happens to you now. There are other cases beyond family protection that one might buy life insurance for example a business might to buy life insurance on a key executive or a super wealthy person may buy life insurance for state planning purposes but for the vast majority of Americans end vast majority of people the reason the buy life insurance is for family protection and as I understand life insurance it is priced through these actuarial tables right like historically. These are tables that basically predict when people are going to die. That's correct correct so actually spend a lot of time looking through population demographics data so on so forth to create mortality tables of every thousand people. How many people do they think is GonNa pass away in a certain year and these tables are created based on the population Galatian of folks in a life insurers historical data it. How good is that data. And how accurate is it in making predictions. That's really good question. Some carriers have more data than others and even if the data is good at the end of the day. A meaningful impact into the performance of that data is how you execute underwriting and underwriting is basically the evaluation. Shen of the Individual Person Against the category of risks that they're supposed to be placed into description. The inefficiencies of the life insurance market life insurance has been around four hundred fifty years. The practices that life insurance companies are using have been very much outdated so for instance the sales of life insurance ninety nine percent of it goes through a life insurance agent the one that you met like the one that I'm at the distribution force is basically human to human and there's very little technology that enables the life insurance companies to actually provided end to end sale online most life insurance companies have dozens if not more of different systems that don't talk to each other things from the underwriting system things from the policy administration system things from the in sales management system these systems don't talk to each other and therefore make it really hard to deliver an end to end seamless online journey for your customers. We'll get into the engineering eventually. I want to make a little bit more context linke. This is your second life insurance. It's company. Why have you started multiple life insurance companies. That's a really good question so after my experience of of buying that Permanent Life Insurance Policy back when I was twenty two I eventually came to realize that that policy that was sold to me was really not the right fit for my financial situation and at that time when I made that realization my co founder Peter Knight decided we will look into this problem. What is essentially we came to realize was that the industry has made life insurance into this very complex morass and we wanted to come in and help. Bring it back to basically the roots of what life insurance meant to be family protection and so we started our first company opted. This is our first year business school and essentially the goal of audited was to help folks who have permanent permanent life insurance policies and sell those policies to institutional investors and that was a way to basically solve a symptom of the problem that I had God which was I had a permanent life insurance policy that I could no longer for to pay and I wanted to figure out a way of recouping some of the money money that paid into it over the past now what we realized with our first startup in life insurance was that that's a limited market and we're solving a symptom and what we really wanted to do with ethos was to solve the root cause of the problem which is helping people get the right kind of coverage in the first place. I I see so the first life insurance business was basically. You've got buyer's remorse. You've got this policy that you're locked into and you want to sell. Will that policy so that essentially instead of you know your family potentially getting your life insurance in the case that you pass away you you basically say I'm just going to cut my losses and just sell this to somebody else. That person that person on the secondary life insurance market would end up making the money and in the event that you pass away correct correct. It was a way for folks to liquidate this large asset that if pay a lot of money into so that's an interesting business to star with in the sense that it is like a limited market. I can imagine the liquidity being pretty low like. I don't know anybody buddy who has well. I guess this kind of thing doesn't come up in daily conversation. I don't know who has resold a life insurance policy or who buys these life insurance policies. Also I imagine at the time you were like. Maybe we could be a category creation or maybe we can increase the liquidity of the market through technology but I can also see it now working out or you just saying well. Let's just go after like the main part of the market like the initial sale. So what did you learn from that at first life. And what else did you other than like the market in a little bit more depth learn from that first life insurance company that you're applying to your current aren't one yeah we learned a whole lot about life insurance in general when we first started off did. Peter Nye had zero life insurance experience agreeance and so there was a massive learning curve in order for us to get up to speed in the market and that taught us a lot of the inefficiencies agencies in the market that everything is sold through humans. It's all personal person relationships. There's very little technology in all the existing technologies really backdated things like underwriting policy administration are super complicated problems that carry yourself by bringing on you know tens hundreds of people to tackle and so that insight was kind of the key insight that led us to eventually start ethos so to get into to the technology side of things visible. Maybe you could start off by just talking about why does building a life insurance company differ from building a traditional Sassou web APP that we might think of or does it differ. Is it just like any typical. Sas sweb yeah that's great cushion while in principle it has not you're trying to build a delightful product which is extremely efficient for your customers for different companies customer for Amazon might be looking for a product and getting it shipped at home as fast as possible for us. The experiences largely the same which is opposed soon is looking to buy life insurance policy and we wanted to get the policy in their hand as fast as we can so the principals and the team structure or the technology that is is required to make that possible is very much the same the problem we are solving higher might differ lot which means that every single day when we're solving problems we're learning more and more about life insurance and we're finding solutions to bring efficiency into that system. Describe your stack. He has searle beating about a stack in terms of the user. Journey person is looking for Life Insurance Co.. Journey starts with okay. What do I do. Where do I go. Where do I buy Life Insurance Insurance. From how much do I need. What am I going to pay from that. Search the end up finding somebody. WHO's ready to explain them? What Life Insurance? This is all about largely. It is an agent in our case. It's an automated product so it starts with what we called acquisition then he goes to a the channel that we call growth which is converting people who are looking to find life insurance into a user that would meaningfully engaged with our product. The next year needs engagement. which is the application submission process which is a very similar experience like turbotax? Most applications submitted. We want to make sure that that policies Z's activated as quickly as possible which is the next journey which includes underwriting and other rotations steps which is what we got activation and then the last step which is administration is making sure the user snow what their policy is they can make changes to it and be cared about retaining for the life of the policy so these are essentially the user journey and the stack that we have is pretty much guarding each part of this journey so some teams heavily focused on understanding user czar behaviors and what they need from qualitative and quantitative perspective so they're very deep data. Some teams are working on providing the best experience to users which which means they're ready heavily thinking about these interactions designed seamless experience that art of things some things are very much thinking about how to bring efficiency into the the stomach is making sure the writing is happening as fast as possible we are able to connect with users in back and forth conversation as efficiently as possible so that's that's generally a stack on the technology front. We are hosted in aws cloud for most of our things and we use all the modern tools that are available to us to actually bring this technology or convert this into a product analytics is an embedded business intelligence tool it allows you to make dashboards and reports embedded in your applications create deploy and constantly improve your analytic applications that engage users and drive revenue. You focus on building the best applications for users while. Lachey gets you there faster and keep you competitive. Logic analytics is used by over eighteen hundred teams including VERIZON CISCO CISCO GO daddy and J. P. Morgan Chase check it out by going to L. O. G. I. Analytics Dot Com Slash Data Science. That's low coachee. Analytics Dot com slash data science logic can be used to maintain your brand while keeping a consistent familiar and branded user is our interface so that your users don't feel like they're out of place. It's an embedded analytics tool. You can extend your application with advanced. Api is you can create custom experiences for all your users and you can deliver a platform. That's tailored to meet specific customer needs and you do all that with Logi analytics Loggia analytics dot com slash data science to find out more and thank you to Loggia analytics. We'll come back to the engineering but I think it's probably worth explaining meaning to the listeners just top down what happens on the website so I went through the on boarding flow. It was certainly friendlier friendlier than interacting with my friend who was trying to sell me life insurance so basically it's a questionnaire. It asked me a set of questions. Go to ethos life say. Oh Yeah I'm looking for life. Insurance asked me this curious set of questions like am I in good health that one made sense but they didn't it ask for like my annual income by Zip Code. I was kind of like why does this matter like. Why does my my zip code. have any bearing bearing on weather get a cheaper and expensive life insurance policy. Yeah that's a great question. So editor of the insurance is about about calculating risk a dash to the policy and that risk is calculated by several pieces of data by going through our application process user user is self reporting data on themselves decorate data is how healthy they are which made a ben whether what's their lifestyle is all about you know what kind of activities they have in their daily life in Vorkuta proficient. They are now sip. Gourd is a great example year the demographic but didn't zip code makes a a lot of difference which means what kind of food you eat walk on activities you are in and that actually is a good indicator of your lifestyle. You're living in that ZIP code. All all of these pieces of information are discrete data points more than you put all of them together. You're essentially trying to predict how healthy an individual will is and based on that got them into a restful which ultimately decides the pricing for the coverage US looking to get. Could you explain the concept of of a risk pool. Yes so in life insurance typically the way you price product is with what's called a risk in class so you might have heard terms like standard preferred preferred best terms like that those those types of terms describe a risk pool that a person might be put into at ethos. We have five different risk pools. There can be more depending on the company and so what happens is when a person comes in to the application their application is evaluated valuated against the criteria of the risk pool and they are put into one of those risk pools and that risk putting it's a major determinant of the price that that your you'll receive as an in customer okay so I think we have a better understanding of the product and we can talk a little bit more about the engineering so I'm getting a picture of this product where basically I go to this web page and so there's this funnel process where you WanNa get me to fill out a questionnaire I fill out the questionnaire. It slots me into a risk pool and then it gives me an offer like I got. I got a couple offers. I got rates for thirty dollars and fifty dollars per month for a twenty year policy. The fifty dollar per year case was six hundred dollars adds up to six hundred dollars per year and then that's twenty years first and that was a million dollars in coverage and so I was doing the calculation and so if I pay six hundred dollars per year for twenty years that's vats twelve thousand dollars for a one million dollar coverage amount and so the implication that I got there. Was that a you. Ethos you would break. Even if about one in one hundred people ended up cashing in the policy keep telling me whether my calculations relations correct and just tell me about how you take the inputs and you end up with a calculation this expected value value calculation and the coverage offer that you give to the user yeah. That's a great question to answer the first part of your question. Yes typically see for a term life insurance product. You hope that relatively few of the term life policies result in claim so that that is is what the actuarial tables that we talked about earlier summarizes they summarize what percent of people in a certain pool over over certain time period is expected to make a claim now chances the second part of your question. How do all of the inputs that we ask for throughout the application process process determine a risk class at the end so we asked a ton of different questions things that you've seen like you know. Do you smoke Moke. What kind of medical conditions might have had in the past. What are some of the activities and hobbies that you do so on so forth all of that information is Dan combined with Third Party information that we collect on you from other. API's data sources and all of that go into an underwriting addie model that then determines what risk class you end up in and you can think of it as a very similar to any machine learning problem where there's set of data in a set of there is a model that is predicting something and then data validation process of the output of that's what we are seeing is that in in traditional underwriting is largely man were process and that's where we see a lot of opportunity and using lot more data and data science to not only designed into better decisions but also the executive as fast as possible. There's so many reasons why this is an insanely good business to get into. I mean I've eight. I don't know a whole lot about Warren Buffett's businesses but I know he likes insurance businesses. I think largely because it's like you know one of the original recurring revenue sources right like people don't really churn from an insurance business that you know assuming they don't get basically taken shaken by some sales person that convinces them of kind of sketchy sales pitch if your car insurance if you've got a car you need home insurance appearance. If you've got a home and it's GonNa be very hard to churn so we know that this is great recurring revenue and then compared to the competitors. I mean gene the competitors are I assume working with you know just Janke. PDF's and like scanned pages and imagining you know these dusty insurance offices with lots of printers and can you just talk about some of the competitive advantages that you can have as it technology company approaching the life insurance business from a total greenfield situation. Yeah that's a great question you know the opportunities in every bar of this final just imagine a person trying to buying life insurance they watch an ad on TV than go and find an agent and then schedule a call with Agen. Maybe on four hundred maybe in person. Dr Them Understand Life Insurance go to this application process which is extremely paper Drouin and would many downs roscoe redundant cushions and cushions that be naughty when applied to a person and then after that a ton done of Muruga dust and blood test for all this data and it is largely fifteen to eighteen week process currently and we are trying to bring that process downing do a couple of days as a company. We are extremely motivated to reduce the time from person starting in my life insurance to the dime that they have within the policy they have in their hand so what is our leverage by using technology when somebody looks at ad on facebook and Google the jump into an application process my really thinking deeply about that application process what question should be asked based on what we know. If you're answering certain questions maybe maybe many other questions don't apply and we may not even actually show you those questions to you. It can be used third party data to bridge the gap so that you don't have to go go to that lengthy application process and removing the amount of time. You're spending and some are filling up their application once you submitted that application Asian. How can we have the underwriting process. which is that a scout collation that we've talked about as fast as possible and if there is any relation or for the data needed for the policy get activated? How can we do it in a modern seamless me like most users are used to more says meeting a person are taking phone calls. Lindy phone calls and things like that so if you look at it that is inefficiency built into every bar this funnel and by using data by using technology apology and great expedience begin significant introduce the time to spend to buy life insurance policy and bring efficiency indigenous and our funnel the other thing I would add to that is one of the major advantages that we have as a business is that we've been able to build an end to end technology stack for all pieces of the life insurance process so distribution which means the customer coming onto our website filling out an online application that is then represented in structure data to be sent to underwriting which we have an engine that evaluate it's that structured data alongside third party information that we receive to policy administration the delivery of your policy the creation of your policy language and document and the automatic managing of billing all of these things are typically different systems and legacy systems at an existing shirts company and so we've had the advantage of building all of these things from the ground up in house so there are all interconnected which allows us to deliver a really positive customer experience. The target is two days. He's to offer the customer a life insurance policy. Yeah we want it to be less than that and that's what we are striving for and Okay Oh so the standard is two days like the industry standard today is today no industry standard currencies fifteen to around fifteen week that why is it so long and why are you only targeting two days kit you get this down to like a hundred milliseconds. That's the goal and hopefully you know as we are our technology stack gas as we understand this problem lot more. We collect data. We wanted to be as fast as it can be and so there's a trade off right between how fast and how much data you have to evaluate someone versus what their ultimate outcome is in price and so you can always trade off collecting less data and making a more loose judgment and giving someone a higher price to account for that uncertainty what we want to offer the best of both worlds what we WANNA do we want to collect data about a person and be able to make a really accurate. Hurston wise decision which allows us to really understand the risk and provide provide a good price to the customer shifting that frontier is non trivial fascinating so in that two shoe day window that you're targeting still understand why does have to be two days like the doesn't comp pretty complex data science workflows get done in in less than two days. Are I guess are there third party data providers that those are just like the longest like I'm wondering what is the longest source of latency in that two day target. Is it some third party data provider. It could be a number of things sometimes third party. Data providers are slept. That is definitely true. Sometimes sometimes we may need a follow up with an applicant to ask them more information because something in their application questions and something in the third party data did not connect or they say opposite things so we have to figure out why that was the case. Sometimes it requires a human to take a look the case to try and understand what that discrepancy is and sometimes we have data from our third party vendors that currently is not machine readable so there are a number of things that contribute to that today's but you're right in the long term goal this should happen under one second if we're able to truly get all of the information that we need in an API structure data format and we are able to process that information in an instant that is along terminal it kind of reminds me of we did an interview with checker awhile ago and Checker is that Api for doing background checks and you know the user makes an API call and then sometimes like the result of that API call involves going to a courthouse to get like you know some information. That's only tailable through going to a courthouse and you said that there's some data that's not human readable so you know the customer makes a request for a life insurance quote and you need to get this what what's an example of non human readable again. I don't want to give away your secret sauce but like what's an example of of a data provider. You have to get that does not provide you machine. Readable information scribbled notes from the physician as an example of cheese. How do you have access to that information though like you know if I make a request to her life insurance policy. Are My our medical notes about my physicians visits. Are there somehow available able to you whom you submit a life insurance application. This is the case or every single insurance company that at least I know up in the United States. You have to sign a hip a form which allows the life insurance company to obtain some of those records from your Whoa Man I bet you you can do a better job at consolidating my medical records than I can. I kick my medical. Records have basically been like lost in the sands of time. It's a move from city to city like you know you go through this as a separate issue but like try to transfer your medical records from one doctor to another is arduous but yeah. I guess you're not tackling that problem today. There was one other insurance show we did fairly recently with a company that that is improving the insurance currence brokerage process and they think that was what is called the insurance brokerage process but anyway they had to interact with a bunch of other players players in the insurance market so there are these multi stage insurance pipeline procedures but it sounds like you're more more on the full stack side. If things are there any legacy providers that you have to interact with like legacy insurance providers absolutely absolutely absolutely so we partner with a number of insurance giants including legal and General for instance one of the carrier partners that we work with and we work with reinsurer's as well so reinsurance companies are insurance companies for insurance companies and we partner with these folks because they have tons of capital and tons of experience and so when we sell a policy for instance our policies are not that by ethos because we're startup up there are backed by a hundred year old insurance company in Reinsurance Company because they've got the longevity to guide the capital base to ensure that our customers are well protected for the twenty thirty years coming forward okay. Let's come back to the engineering side of things so how much quote unquote quote data science is going on today and tell me about your data engineering pipeline full yeah absolutely so be v used data for many things improving our product digging better decisions understanding customer behaviors and also understanding has defar company the organized idea deems is into different parts one that is purely amardeep engineering which includes the by blind essentially instrumental in every part of our APP to understand how users are using application also the data that we get get from third parties as well as all the data that we have in our databases bringing all of those different pieces of data at one place so let's begin to data mining deeper insights and radius other things with that so that's the PR data pipeline data warehousing part of things then the machine learning running back which is essentially using his data to fundamentally improve the experience for users within the product and lots of things that we do for example right sizing thing is one which you wouldn't find a traditional insurance company doing Rickett about our customers that they're buying the right size policies so that they can afford it for a long period of time so we used data science to understand that it about a user and suggests them the policy in the coverage that they should buy and that's data redesigns problem. There are many many examples of such things that we do not educate our customers but also make that experience better all of these things which which is the data pipeline data warehousing machine. Learning is what engineering is focused on then the second big bottle of this problem is really understanding in how what users are using a product which means deep understanding of this. Dr Funnel like how they're progressing on every single page every single interaction action in are we seeing the biggest friction mayor we see the biggest drop offs in those funnels read on a lot of ab experiments due improve the experience of our product. How are those. Ab Test Experiments Doing Diba standing of those all of this is what we got produc like signs and our product team and product managers are deeply inboard into this aspect of really deeply understanding user behavior and then comes the energetic spa spa which is also the strategy confident which is at on pricing how you're spending requiring customers what is the efficiency of each channel all of those things that are very strategic piece which is in our business operations team so we see data not only solving big parts of the product but Berkshire just make everything inside the company and every single decision making as efficient as we can right so basically we talk about the data side of things saying to new problems that are related but probably use fairly different stacks so one problem is that the on boarding flow hello data problem in the sense that if you're trying to get people to fill out a form and make it through a funnel for purchasing being a life insurance policy and then there's all these things around like do you send email follow up stu. You show Google ads to people. This is one side of the data problem. It's kind of the architecture of your funnel. That's one very interesting problem the other very interesting problem. Is this data data engineering machine learning problem where you've got tons of data. You've got a user with a certain set of parameters and and you want to match that user to the right sized available policy or policies that you can present them with so. I'd Kinda like to talk about the stack. We talk about each of those problems independently but I'm just very curious about the stack because I've talked to so many companies that are building in a quote unquote data platform that involves an et L. process a data lake a data warehouse US several different databases streaming frameworks a machine learning framework like tensor flow a bunch of random scripts written in python eighth on. You've got all these different things fit together in a quote unquote data platform and there's there's not really like best practices around this kind of thing. It's SORTA like people just duct tape and chicken wire these different pipelines together because that's kind of a state of the art to the extent that I understand it. Tell me about the engineering genera decisions. You've made around building your data platform yeah exactly so you should many great things and these are the problems that almost just every company's trying to solve the duals might be different but the principle stays the same for example. We would like to understand how our users are using product product beaches. They are interacting the most bit what burdens clicking the most walk part of the experience that they stop or close the browser and do not want to engage the park anymore. These are all things that we can capture by instrumental in our application in the right way then the third party data and the story in our database you have to bring all of these things together so that you can make deep insights on it because if a piece of data sitting in one database and another piece of data sitting in a second database it's very hard to join across those two things to get deeper insights and that's essentially what the data pipeline and our data warehousing problems so worn is our stack be usable stress for our main database. We rely on of Amazon tools which are out of the box largely. We're still early in our journey as a company so data. Walloon Walloon is in a significant problem for us more. We are taking decisions on as the efficiency and the speed of execution so we rely on lot of third party tools do bring this data into the data warehouse use redshift and the ones that Dana is in wretched. We have an email process which is larger than by airflow where we can actually actually this piece of data and transform it into different tables. Are we use on which we can drive. Incites ready easily now wants these tables tables and these views already be exposed them in radio space on machine. Learning applications could be using this data to genetic features to train the predictive modeling and Jimmy Outcomes offered our our product science team might be generating funnels on topic to understand house a user going from one base to another beach and how how much time they took and is the performance right and where the drop off sports and machine learning stack is largely driven by by Tom. We be still not using deep learning because we are very problems are extremely related to regression problems and problems that odd more predictive in nature and so be it using things like logistic regressions and few other algorithms using some really amazing. Biden toolkit and we drive funnels and charts using lot of third party tools such as he CIARDI. Oh and things like that prosper ops is a platform for automatically managing reserved instances on aws reserved instances or one of the most powerful tools to to reduce your cloud bill where aws gives you discounts of up to seventy five percent in exchange for a term commitment you don't have to change your APP or aws the US environment to use reserved instances so your savings start immediately however using reserved instances can be both complicated to figure out and time consuming to manage prosper ops simplifies that complexity by using algorithms to build and manage a reserved instance portfolio that maximizes your savings even as your e c two footprint changes go to software engineering daily dot Com slash prosper ops and get a free reserved instance stints savings analysis prosper ops can more than pay for itself with the savings that it can generate so checkout software engineering daily. Dot Com. Tom Slash Prosper ops and save time and money with prosper. OPS reserved instance management. Thank you to prosper ops for being a sponsor of software engineering daily really shows. You're starting to get into this machine. Learning driven approach have you noticed any pricing differences between what your machine learning models suggest suggest and what the insurance giant's recommend traditionally. I think we have seen some of that. Ah But we're very careful to make judgment calls on that at our current stage because of the limited amount of data and experience we have we in conjunction with our insurance partner spent a lot of time looking at not only our data but also historical data and over one hundred years of experience on life insurance mortality so for sure but we are still very much in the learning phase and exploration of that I would retreat what linkages said a little while ago that pricing and how much do they collect are two things that we have to balance against each other because it's essentially a risk allegation. If E- have abusers spent a lot of time application process maybe the underwriting process would be faster which desert into a cheaper price but we might give them an experience which might not be really efficient because we are taking them to lend the application process so it's a balance that we have to keep in mind to make sure the a user are deeply engaged with the product but also get the right price for what the needs are okay interesting so like let's go a little bit deeper on and a specific machine learning model that you might be building. Can you just give me an example of the end to end workflow for bill. Maybe give me an example model that you WanNa talk about just the process of getting the data. Maybe cleaning the data putting the data into up database or data lake or you know in memory system. Whatever just give me the tell me how the sausage is made yeah so I'd only talked about right sizing. They choose another problem just to bring diversity into this discussion so churn Johnny something that we deeply care about because we want to make sure that users are with us for the entire life cycle of the author policy and shown is essentially a predictive modeling problem when people are some meeting an application when they are getting the policy activated. We want to make sure that everything is going right in order for them to stay along with their policy so so now if you want to predict the churn. Let's say we want to predict whether somebody who is engaged with thousand got an activated with journal not the data of the applications that the food would-be not protests database we will take that data and the toad body did other we have on them and bring it into our data warehouse now. Our data warehouse has this entire doc set of data that based on their application. Their behaviors are they using the product and all the third party data in one place now machine learning problem is essentially saying given what we know about this user botched the probability of this user churning in two months three months six months eight months ten months into predictive problem and so we look into ways of the signals that will just like a regular machine learning models we come up with initial set of features or attributes that we think about Imboden Auden to make this prediction and then we train it based on the data for example set of people that have not shown and a set of people that have shown and on machine learning models would not see those two things and stock the way these attributes based on the data. They're seeing and will result into some kind of model which will have fought each one of these attributes now in a user new user comes in we can take them through the same pipeline where we have all of this data for this new user and we take those attributes can they do those attributes and based on the model that we have a while ago cannot predict whether this will this user churn and what would be the key attributes why charn when we for example one of the things that were discovered when we look at shown is that one of the big reason that people churn learned is that credit card expires and forget to make payments now by being a technology companies easy for us to Saul is build a notification systems stones into the pipelines so that we can now remind us that hey your credit card is about to expire and you might actually shown lose the policy. Please go ahead and update it by using machine learning now. We have solid critical part of this journey problem. Let's zoom in on one particular aspect to the data warehousing the definition of data warehousing using as I understand it. The kind of the modern definition is if you want to get a bunch of data into a place where that data can be accessed. I really really quickly you often put it into a data warehouse so data warehouses useful for performing a lot of calculations over for large data sets and it's expensive because I guess this data is is usually in memory or it's in a place where you can access it faster astor and so you can't just keep all of your data in a data warehouse. Can you tell me how a data warehouse fits into your your engineering process. You Know Windy. You put data into the warehouse. When do you take it out of the data warehouse and what systems are performing those those. I think that's called the ETL where you're putting data into the data warehouse and taking it out yeah absolutely so the reason why anybody would use data warehouse. It has many things to do like one of the things they're deductible on the efficiency of the query. When I run a query how fast it can run without a data warehouse. If the raiders in memory is is one aspect of it but imagine if your data is stored into three separate databases given every company's stock would use multiple data stores doors to store that data they might be starting flat files toward body information on S. three they might be starting structure delayed a relational database like my sequel or post-chris they might be using a no sequel system where they want to get high reads outburst by writing fast something like the Sandra or Mongo. DB Now the since the data is spread into these different locations when you need a query where you need to join a girls these three different things it's extremely hard because now you have to to figure out a way to connect the dots between three different stores. That's what data warehouses becomes really handy because now you can eat all of this data and bring it into one central place now. Normally you have you're saving time on on were also all of this data sits very close to each other and you can structure today no way eh whether if it isn't different tables then you can join across those staples and by using modern systems like prester big quit you can do it extremely fast all all you can go to an et. L. Layer and create a view or a table where you bring the data that is needed for query into one place so that you can quitting on a fast so data warehouse not only solves the speed problem by exhibiting equity fast but also made it extremely accessible by bringing the different pieces of data in one place and extremely easy to access that data how do machine learning frameworks and streaming streaming frameworks. How do these things interface with the data warehouse or or to what extent do they interface with a data warehouse. Yeah so innovative simply data warehouse is kind of your large database in which you can store your indata said in different views and machine learning problems. You need to be able to get that data run. Your models trainer models bring being output from the status now. Some of these problems are batch processes for example. You don't really need the real time decision making in order order to make decisions so for example like for our churn if Iran it once a day that would be fine for us but if there is any other problem problem for example if you wanted to reflect in via time any bar to our customers on our product may be goats or some other thing that act might need a real time machine learning model and that's where it streaming comes into place which can stream data to a model in real time we can use the smarter to make decisions and streaming back to our product so it's largely between batch process and real time processing and different systems. Do those things okay interesting. Let's zoom out a little bit crunch. Base tells me. Your company is fifty to one hundred people. Tell me about the experience of Scaling Ethos yes. That's accurate. Were at about eighty people right now two years ago. I think we were maybe less than ten people and so it's it's been it's it's been very exciting. The ability to scale the company that also comes with the number challenges to be totally candid Peter in IR managing such a large team team for the first time and were very fortunate to have brought on exceptional experience leaders like people who has managed a large team in the past and so growing team is Super Fun because now we're able to tackle bigger and larger problems now when it was just you know eight of us and so that's been super exciting and something that were looking forward to grow in the future. How's the company structured at this point in the sense of. What are the different teams because you know you've got these interesting data science problems? You said you have a a role. What was it near note scientists scientists. What is a product scientist so product scientists are folks that have data expedients but also the product expedients particular data scientists ordeal engineers were largely look at the problem from engineering point of view they will be ready excited in the scale of the data the processing of the NATO also building models in a way that they can derive some are are solved problem but product sinus are skilled in understanding that data and relating it to how to make changes to the product and that's a very product function so head of product Google have deep experience? It's coming from facebook instagram snapchat and one of the things that these companies have done really well is used product scientists to use the data that the engineering getting is bullying and make the product constantly better by learning from and that's what Barak signed is are what they worry about day to day. Basis is how is our product Funnel Donald what stages odd the funnel what attributes and and thinks we should be looking at is an experiment doing better on desktop words as mobile as a political expedience experiment doing better in certain segment of users was is not certain segment of users and these deep insights are extremely useful for us to continuously. Ashley cleared and experience that I use and it's not just quantitative because there's a big quartered anybody's to it which has made design team plays a big role in constantly talking talking to our customers running user testing and also keeping understanding how users are are warranted talking about a product and take this quantitative and qualitative data that we gather from Betas sources and make data informed decision and that's what product is all about and then to go back to your earlier question question of how our company is structured in general we have different functional areas so we have engineering product design the customer experience marketing legal so on so forth and there are also cross functional pods that worked together so for for instance our growth pod includes not only engineering and product and design but also includes our customer service reps so on on so forth and our analytics teams and so that's how we think about the company in slices of the customer experience and that's a major factor in how we structure our teens a VIP UIL. What's the hardest engineering problem you've had to solve in working at ethos started audio engineering team was six people and now we are getting closer to thirty and to me when I think about problems had taken about three ways which is product people and process and my role is to think about all of those three things at the same time which as we are scaling our engineering team and hopefully from thirty to sixty ready soon what processes are working and what processes are not working a certain set of agile frameworks which might log Ackman your den will will not work when you are at fifty and house our culture evolving along with that so that's the process part of it the people part of it. What kind of people do we need. A different different stage of the company are part. You're extremely motivated to ship product as fast as you can little bit later part you want to do that but also ensure that the quality ladies high and you can ship fast which brings a lot of architecture problems into the plate so currently we are building an architecture team which will take aac on application and provide different services on top of it so that different teams can big ownership of their part and move really really fast so the current problem taking his APP is GonNa breaking down into different. Services is a very interesting problem to solve and then comes the product which is thinking about art like now this application process and it looks very simple when users interact with it but if you look at behind the scene it's a very broad and deep treat st because based on the answers you're giving to the cushions your next set of questions will appear and this could be extremely slow process to traverse that tree and we. WanNa make sure that that is as fast as possible so we're constantly thinking about. How do we give this expense to the users in really really fast and that's a very interesting using problem. That team is working on right now all right last question linke. Tell me something non obvious about the life insurance business business that you would not know if you had not created to businesses in this space. I think that's something I wouldn't have known without. Having dug deep into this industry is actually how much technology can change what we do in this industry. I think it's easy to say from an outside perspective. Hey if we build brand spanking new systems that it will the make things more efficient but I think that's a very surface level way of looking at it at the end of the day. What I've come to realize is that yes building a end to end system from the ground up is super important but to it's really the data and it's really understanding of our data and our model also will allow us to deliver a phenomenal customer experience at the end of the day and really what that takes redesigning the end to end stack walk of life insurance from a tax standpoint from a process standpoint from an experience standpoint from the ground up in order to succeed okay. Hey guys will thank you for coming on software engineer daily really fun talking to you. Thanks so much krisha opportunity. Thanks Jeff. Thanks for having us on take software engineering daily reaches thirty thousand engineers every weekday and two hundred and fifty thousand engineers every month. If you'd like to sponsor software engineer daily send us an e mail sponsor at software software engineering daily. Dot Com reaching developers and technical audiences is not easy and we've spent the last four years developing being a trusted relationship with our audience. We don't accept every advertiser because we work closely with our advertisers and we make sure that the product is something that can be useful useful to our listeners. Developers are always looking to save time and money and developers are happy to purchase products that fulfill this goal. You can send Edison email at sponsor at software engineering daily Dot Com. Even if you're just curious about sponsorships you can feel free to send us an email. We have a variety of sponsorship mantra ship packages and options. Thanks for listening

Coming up next