Placeholder Image

字幕表 動画を再生する

  • and earlier this year we just sold Thio polio and which is where I am building out the product, scaling it on DDE, adapting the product to their client base.

  • Got it?

  • Mind if I ask how much did you get?

  • I sell it for Well, it's public for people who want Thio.

  • No mind if I do my research and put in my video?

  • I guess I guess I can just tell you because you're gonna put that in any way Way sold for 60 million Before we continue this video, I just want to say thank you, Brilliant for sponsoring this video Everyday Brilliant publishes daily challenges on many stem topics like math, science and computer science.

  • This site is extremely sleek, and they have over 60 interactive courses, which makes learning these concepts way easier because of the hands on approach.

  • They also have a artificial neural networks course, which is really, really well made, and I think if you're very interested in ML, you should definitely check it out.

  • This is a great compliment to university because you can do practice questions under well curated sequences of problems, which allows you to master the topic.

  • You want a master.

  • You know, I wish I had brilliant when I was in college because I'm more of a hands on cam guy.

  • I learned by practicing, and if I had brilliant, I think I would have understood these concepts.

  • Ah, lot faster.

  • So if you're interested, you can go on brilliant dot org's slash Joma and then the 1st 200 people, we'll get 20% off.

  • All right, that's it.

  • All right.

  • Welcome back.

  • Thank you.

  • Thank you.

  • Thanks for having me back.

  • Thanks for coming, David.

  • Ma light brother calls, I think your first interview that came back for a second round of my right.

  • I think you're actually Yeah.

  • Interesting.

  • Because you were high in demand, so I had to bring you back a cool.

  • So just a little context.

  • In the previous video, I made an interview with you, and it's mostly about how you were a quantity at Two Sigma.

  • And then that video you told me that you quit your job to do Dicko found a startup that correct.

  • So a lot of people were wondering Why did you make the switch the switch?

  • There's a lot of smaller reasons I wanted to try something new wanting toe.

  • See what was out there for a deep learning.

  • I wanted to also look at crypto currencies.

  • I want to look at biotech.

  • But Indian Elliot, co founder of Dynasty Re shot me like a week after I said I was gonna quit.

  • And he was like, Hey, have you ever thought about quitting?

  • And I was like, I quit, like, one week of ago.

  • How do you How did you How'd you do this on?

  • And then you said, like, Yeah, I just come out to l a and then see what we're doing.

  • And then, uh, there's nothing no harm done, right?

  • I know one thing led to next and, you know, here I am now.

  • Nice.

  • Nice.

  • So, between back then, when I interviewed you, too now, any updates on the startup is dynasty, right?

  • Yeah.

  • There was a lot of a lot happened since since the last time we spoke.

  • Um, I joined Dynasty.

  • At that time, we had just pivoted into, um, the the A I for real estate business.

  • And during that time, way built the product, found product, market fits, um, and earlier this year, we just sold Thio polio and which is where I am building out the product, scaling it and adapting the product to their client base.

  • Got it?

  • Mind if I ask how much did you get?

  • I sell it for?

  • Well, it's public for people who want Thio.

  • No mind if I do my research and put in my video?

  • I guess I guess I can just tell you because you're gonna put it on anyway.

  • Way sold for 60 million.

  • Nice.

  • How many co founders are you guys?

  • So Dynasty was initially a different business which did not succeed on.

  • That's very that's very common for a lot of startups.

  • Esso E At the end of the last business, everything was were like 10 people on, and I joined that as like the sixth person.

  • At that time, things were not going well.

  • There was no product market fits and we started pivoting into the real estate and about half the people left.

  • So in terms of co founder, originally there were two co founders and then this'll new will for the new pivot.

  • Five of us were left just a t l D r.

  • What was the previous product and what is the product now?

  • Just to make it more clear.

  • Back then, we wanted to create willing to securitize real estate assets, basically create a an exchange and allowing people with a smaller amount of capital to taken take positions in real estate assets.

  • If you think about it right now, to buy a house, you have to have, you know, especially in Silicon Valley, you have to have, like, 200 K 300 k just to put it put down a deposit.

  • It's not very Democratic on also, your all your money is into it like this one single asset that's very susceptible.

  • Thio, um, to local changes, eh?

  • So we want to change that.

  • It didn't really work out.

  • Well, at least we didn't find how to make it work.

  • I'm not going to say it's a bad idea because we still think that, you know, there are benefits to this world we're dreaming of.

  • But in the process, Elliot's and other people who joined before May found out that a lot of real estate participants had a lot of trouble managing their assets.

  • So it's unlike stocks.

  • Real estate is, ah, an asset that you have toe, you know, it's a real thing you have thio keep.

  • Yeah, there's up.

  • Gave you wantto get people into for rentals and stuff like that.

  • I think half of the income comes from rentals if you take the other half as appreciation.

  • So one of the big challenges was the operations of leasing a building or releasing your assets.

  • So we decided, Well, everybody says it's a problem.

  • So let's do something about it.

  • Yeah, that that's where Lisa came in, he says.

  • The second iteration of Dynasty.

  • So that's like the pivot That's your new business.

  • And of course, Lisa is an A I for leasing the I was wondering now that you dig your hands deep into ML order, any misconceptions about ML engineering that you want to debunk?

  • I think the general excitement about ML is great.

  • It made a lot of people going to Malibu, and that's awesome.

  • But like a lot of focus has been on.

  • How do I build models and how do I fit fit a model t o data, but like very little focus has been on.

  • How do I generate data?

  • How do I design a business process that will create data for the algorithm that I want to build.

  • How do I handle the outputs?

  • How do I build all the process around three ML components that there's too much focus on building the models, not in the focus on how to integrate ML into existing products.

  • And to be fair, it's kind of a new field, right?

  • Not many people have to know how to do this because it's it's It's so new, like analogy is when you know computers were first invented.

  • Her, like the Internet, was missing, that people were finding out how to integrate that into existing business processes.

  • And, you know, it took a lot of trial and error.

  • But that's the same thing.

  • Like not everything is just building models.

  • But not everything becomes more useful if there's machine learning in it.

  • Maybe not.

  • Everything should use Blockchain technology exactly right.

  • That's what that's the example to use.

  • Like, always wondered.

  • A lot of people want to do machine learning now, my viewers, especially because I think it's because I'm in the intersection of data science and soft engineering.

  • But, um, I don't I don't really understand the appeal of machine learning at the job because in my mind, what you do is like you said, You make sure you have good data.

  • Make sure you solve a problem with your email.

  • So most of the time in my head you build data pipelines to funnel it into your model.

  • You pick a model.

  • You play with the you tuned the parameters and then try to optimize for that, you see, and then that's it, like is or is there more to it?

  • I don't know that it's fun.

  • I think like they're they're kind of right.

  • I mean, especially if you're building products.

  • You know you don't have the time.

  • Thio do the fun stuff in animal research.

  • The way I see it suffer engineering is the core skill, and then there's like ML engineering that that gives you a bit more domain knowledge into how to build products are male models.

  • But Indian?

  • One suit.

  • Once you've done that like a V one of it, that's it.

  • You have to have toe build all of the systems around it on.

  • That's not what the school sound, what, you really want it at school.

  • So imagine the fun stuff that machine learning researchers do.

  • What is that?

  • What is the funds?

  • What would be fun for me?

  • I think in terms of research, would be investigating the latest algorithms and like understanding why they work, why they don't trying to fence that bit of sets on these new algorithms.

  • A lot of the things that you've seen out there like Ganz, like generative adversarial networks you make like funky images.

  • Style transfers like these were all investigations.

  • And why do convolution all networks work as they do?

  • And that's the research, right?

  • And that's not it's not primarily those.

  • Those things were not built primarily for business.

  • God.

  • And of course, Lisa is an aye aye for leasing.

  • The one of the problems that people had was like When you put your, uh, your apartment for on Zillow or something like that, you get a ton of inbound, you have e mails, you have text messages and phone calls like, Yes, it's very fragmented, the things that you get and you have to take all of that inbound coordinate showings, take care of applications, and I threw all that moving process.

  • So, Lisa, what we decided to do was thio automate the responses, text messages, e mails, phone calls and on the other side, we just produces showing people just had to show up and at least the leasing agents just have to show up and sell, sell the property.

  • They didn't have to coordinate and do all that stuff on.

  • That was the main, the main focus.

  • I think a lot of people liked how you know, their phones stopped ringing after they used They started using a buddy so cool.

  • Awesome.

  • So what about you?

  • What did you do?

  • Dynasty?

  • You worked on me, So I guess, Yeah, of course.

  • Uh, So, originally I was hired for to do some research because, you know, I was a researcher, blah, blah, blah, but, you know, nothing turns out as we plan it and I decided tow, take that opportunity to, uh, dive into M L.

  • I mean, I had some ML experiences back in college and a little bit back in my previous job, but never deeply.

  • So I decided to build out the ML components, got into deep learning, learn about NLP, which is natural language processing.

  • And once that core thing was built, you know, we were still start up.

  • We just had to do other things.

  • So I I got myself into a cell for engineering, you know, before that had never done, like, really suffer engineering.

  • But under our CTO, I was able to I learned a lot about how to build, uh, two products and like, yes, I did suffer engineering and product design.

  • You basically have to do everything that I started so white.

  • Why couldn't other people just do the same thing to supply a I to real estate stuff?

  • And then would they be a competitive er, like how?

  • Why what made dynasty successful?

  • We also thought that there was gonna be competitors, and I think there are.

  • But nobody took the similar exact approach that we did, which was Lisa will talk to two prospects prospects being the people who would rent, they were protected them without telling them that it's about anything, because in reality, we're not fully about either.

  • About 40% of our messages are handled by humans who who we call operators.

  • The fact that we try to give such a natural, experienced prospect is something that declines really liked because in general, people don't like to interact with the body.

  • So our conversations look very natural because there's a lot of humans that kind of, uh that we fall back thio when things go wrong.

  • Actually, I have a funny story about people not liking to interact with butts.

  • So one of the one thing that we measured was how often with people reply when we send messages back to them.

  • Right is like reply ratio, whatever.

  • Initially, we would reply instantaneously because people were message in that were like, Hey, we have a showing at 2 30 you want to come within seconds?

  • And people were probably freaked out and replied less than if we were to wait two minutes before replying.

  • So although we can reply very quickly, sometimes we wait a minute or two before doing so.

  • That's interesting.

  • So another part that made us successful, I think, was the fact.

  • Well, I think the biggest, the biggest success was the market validation that was done prior to the pivot.

  • But that aside, another thing was our willingness to just build what was needed, like not focus on technology too much and just get something out out there, get it in front of people illiterate and billed it as simply as possible so that we don't waste too much energy trying, tow, optimize a system that was not going to exist like a week later And also these days, eh, I products are like very, very hot.

  • A lot of people spend a lot of time optimizing that small, small component which it, in retrospect, like I could have spent more time on that.

  • But if I did that, I wouldn't have used my time Thio build out the product in terms of the suffer engineering part right, Deciding Thio say Okay, that's good enough and work on, uh, like the most important part for the business was something was was spared that like everybody had Afghanistan.

  • I think that really, you know, pushed us beyond the edge.

  • Gave me a concrete example of that thing you just describe what could have you optimized on like, how did your email system work as a whole?

  • What was that part of the ml system that you could have optimized on?

  • But what did you work on instead?

  • To make your product better in the back end?

  • The first ml component that we created waas intent to classify him.

  • So we would.

  • Taking messages and understand was the intent to classify them as, uh, one of a few intense like Do they want showing did accept something like, Whatever I'm actually before we continue.

  • Can just some kind of have a high level overview of what Lisa does?

  • Because right now I think it's kind of just like a chat, but for the sink, and then is there anything that does?

  • How does it interact with the the operators and also the clients?

  • Right.

  • So Lisa is a pretty complicated product in terms of how you would explain it, like with components that you understand because there's a lot of interactions.

  • One interaction is with the prospect eso With through there we it's just a bunch of texts or emails.

  • The first text comes in, we asked him, Do they want to showing?

  • Then the conversation can go anywhere.

  • They can ask questions about the property they can't like, you know, decide, not Thio.

  • Do anything.

  • We have to confirm them for the showing, and I we ask them how how it's going.

  • That's one interaction with the agents, the leasing agents.

  • We schedule things on their calendar.

  • We asked them questions that we don't know the answer of.

  • We answer questions they they have for us.

  • And there's like an interaction with bosses.

  • Well, which bosses?

  • The boss is a real estate agent boss, sometimes the same person, but often it's not.

  • They want reporting, so that's, uh, best in Maine interaction.

  • And then there's our human operators, which uses Lisa as like an interface to the outside world.

  • They have what we call the command center, and it's mostly like a messenger interface, like augmented with a lot of information.

  • And also they have a concept called the Quick Action, which they can quickly find a commonly use action to reply to prospects.

  • Awesome.

  • Okay, so that means back then, real estate agent will have to communicate with the prospects.

  • Do the showing scheduled?

  • Your own showing managed her own calendar, but now they actually don't need to talk to the prospects.

  • And you just happen, he said, that it's kind of like a layer in between them.

  • That's really cool until they get to the showing, which I don't want to use their human specialty due to sell Well, I mean one day you can build robots and just replace that too.

  • Cool.

  • Okay, so now that we have a good overview, let's go back to, um, what are the M o parts diver specific?

  • I'm pretty sure there's ML in the quick reply.

  • There's probably Ml in trying to classify what messages are.

  • So what were the other stuff that you could that you worked on That was better for the business.

  • And then just optimizing once we got the intent classifier