music]

0:07

[music] All right. Well, welcome back to Data Day with Greg Michaelelsson. I'm joined

0:14

by a great friend of mine, Tadre Sangupta. Uh we met years and years ago at Data Robot. Uh welcome.

0:21

Thank you so much, Greg. Thanks for having me. Yeah, it's my pleasure. It's really good to have you. Uh so we haven't actually

0:27

chatted in a long time. it'd be great to catch up and and stuff like that. Why don't you start out by kind of giving us a little bit of background and kind of

0:33

how you got to where you are today and and then we'll talk a bit about what you're up to and and so on.

0:40

Yeah. Wonderful. So I think uh we met in 2015. So I uh started talking to data

0:46

robot end of 2014 and by the time I joined I think it was Q1 or Q2 of 15 and

0:52

you were already there. We were the first batch of what Jeremy and Tom called uh customerf facing data

0:57

scientist and uh I think these days people call it forward deployed engineer and things like that. People would be

1:03

working with customers and clients and embed their solution into the companies. Uh so before joining data robot I come

1:09

from a data science background. I studied mathematics statistics in school originally from India. came here in 2004

1:16

and started my career with larger companies like uh was very fortunate to start at Deoid uh and then Liberty

1:22

Mutual and then uh AIG New York City. At AIG we got to work with uh amazing folks

1:28

and I kind of knew Jeremy uh when he was at Travelers. My wife was an intern

1:34

under him and loved it. And then I talked to him when I was uh also moving to AIG. had offered from travelers and

1:41

then at AIG I was working with Owen Zang was the top data scientist at Kaggle at that time. So we had incredible folks

1:47

over there. Then Wayne joined data robot and my very last project at AIG was

1:52

automated machine learning. So we I had to build models for 200 underwriting divisions at uh like you know commercial

2:00

insurance part of AIG and uh so put together uh four computers and each of

2:05

them had four cores and ran just one algorithm GBM and uh just uh did a

2:11

parallel group using R and uh like you know kind of was able to so all those 200 underwriting debution has different

2:17

data set that was the challenge but did not really change different algorithm and but

2:22

what were the models what were you predicting So we are beting the like you know which are the more valuable insurance

2:28

submissions. So submission prioritization. So AIG is one of the largest underwriters in the world in

2:33

commercial insurance and they get lot of insurance proposals from the brokers of like AON and other and they just don't

2:40

know which one they should underwrite. So but they do have the data that will tell them that you know historically you

2:45

underwrote this ones and this one was good and this one was bad like you know. So using that data we can predict where

2:51

they should where their underwriters should put their time and that's known as summation prioritization model and uh

2:58

and and and like there are different other proposals that you can use a mixed model and non-machine learning models to

3:03

do the same kind of things like you know how to build the 200 underwriting division is the bigger challenge. We're

3:08

not just building one model for on one data set. We have like you know 20 200

3:14

different data generating process. So how you can model that right so that worked out and uh truth be told I did

3:20

not actually understand what Jeremy they were doing before that uh but once I build that project like you know I

3:26

understood oh my god if I can do this one model and this do parallel and everything it will be everywhere like

3:33

you know so that is when I actually truly appreciated what data robot was doing. reached out to Jeremy and then I

3:39

think reached out to Wayne first and then Jeremy and uh and talked to everyone in December. Razi was the first

3:44

person uh Razi is one of the earliest at data report and Sergey and uh like you know our common usual suspect. So I

3:51

think when we joined it was still less than pretty much less than when I was in tribute was less than 25 and by the time

3:56

I joined I think it was still less than 40 50 people. Uh one of my highest uh I

4:02

had I worked at uh like in Wall Street AIG was in Wall Street, Liberty Mutual was in the financial district of Boston.

4:08

Um but my still favorite memory is Chhattam Street office of uh Boston like you know data robot right like know

4:15

right I think feno market or something like the downstairs was the bar and the middle floor was the law office and the

4:21

third floor was all of us and people are hitting everyone like you know because we were so like you know kind of close

4:26

to crack diet and uh like you know so it's like unbelievable memory and uh yeah so

4:32

that's pretty much uh data robot taught me three things so I I already learned uh at Fortune 100 company how to

4:38

navigate in a complex organization make things happen in a complex thousand 60,000 people organization right but

4:44

data robot taught me three things the first one is that what is a startup and what is the meaning of building under

4:50

tremendous uncertainty I was scared like you know that no one will be closing anything in the first month you remember

4:56

that break very well and like you know we'll be running after the same accounts like we got one account and they're not

5:02

understanding what we're doing a IML was not everywhere at that time so learned really that what is building in uh as a

5:09

startup. Uh tremendous learning and then the second thing was like really horizontal uh exposure. I was always in

5:15

insurance. I was always in a data scientist. So uh within insurance I got lot of exposure already commercial

5:22

consumer etc. But never really got to see healthcare problem, banking problem or working with sales team and

5:28

tremendous learning just working with different people uh and uh people coming from different thinking that was

5:35

tremendously helpful. And the third thing is that like you know kind of really uh data robot was very

5:40

crossunctional. We had of course our functional focus but we'll be working with marketing we'll be working with engineers we'll be working with

5:46

different kind of customers that was that diversity across like whatever we are doing was very enriching. So Jeremy

5:52

Albin knew when I joined I told him that uh I want to start a company and Jeremy

5:58

told me yeah you'll be uh like you can do that very easily and he was right and that training ground was really helpful.

6:04

Four years later I just start started my like you know company I jumped into it at that time it was Hel's insurance the

6:11

original idea that I thought about which became first forward uh how we build infrastructure for home services

6:17

industry uh home service industry as a uh like you know kind of a mathematics we call open problem. So home service

6:23

industry as a open problem. Homeowners struggle all the time how to find highquality providers and how to get a

6:29

job done with reliability. That means 100. Wait, wait, wait. You're going to have to give us some more detail. Uh,

6:35

providers of what? So, home services industry is like one of the largest industry in this country.

6:41

So, American homes is I think second largest asset class, $46 trillion asset value. We love our home and if we own a

6:49

home, we have to really uh like know do many different things in our home. It

6:54

could be a small handyman work uh or put Christmas light on the roof or like you

6:59

know it could be HVAC servicing or HVAC just stopped working. You cannot predict

7:04

when you need something to run your home. Some of us are more handy, some of us are less handy but even if you're

7:11

handy, you need to spend around one to two% of your home hand every year on

7:17

repairing and maintaining the home, right? And that entire marketplace is very inefficient uh for two reasons.

7:24

Number one is that it's very difficult for homeowners to find highquality pros.

7:29

In the next 2 minutes, 3 minutes we can find any pros, any plumber, any handyman, any HVAC that's there like we

7:35

can find them, but there is no guarantee they are high quality, right? So, and then the second thing is a more

7:40

interesting problem. Uh Greg, you also have a statistical background like me. Uh the process variance of home service

7:46

is very high. What do I mean by that? when we go to Starbucks and ask for a coffee uh like like there is not much

7:55

process variance like there is a very systematic process of making up like you know kind of Americano and they'll make it they will give it to us and that

8:01

process is very low process variance cannot go wrong home service extremely high process

8:07

variance like it can go wrong in many different ways like you know kind of it's extremely high and we have seen

8:12

again and again when you bring automation automation has this magical like you know aspect that we all love to

8:18

celebrate But automation is also very boring. It kind of minimize the process variance in a systematic way like

8:24

parallel parking. what Tesla is doing with parallel parking like because it kind of streamline everything and

8:29

simplify and eliminate all those complex nuances and that was needed to that needed to happen in the home service

8:35

industry and that is what we have been doing like you know really wrote down like you know kind of what is the

8:41

meaning of fivestar quality service in home services industry and let's write it down write it down for the homeowners

8:47

and write it down for service providers because we have 6 million independent service providers in this country they

8:53

are very good in what they do in terms of the handyman but their back office or the communication or running a quotation

9:00

running an invoice or sending a payment everything is very very old school and and very unsistatic. So we are basically

9:07

building the infrastructure to power them and as a return homeowners are also getting amazingly delightful experience

9:14

in running the home services. So again wrapping this up two problems on the homeowner side that is how do I find and

9:21

know that I'm working with a high quality pros we are making that happen and then once I find a pro once I say

9:26

okay I'll work with Hector or work with angel how you get the job done with like

9:32

reliability and delightful experience for service provider side in the market we are making sure that their homeowners

9:38

can book them rebook them refer them anytime so they don't have to worry about growing their business but also they're saving a lot of time thanks to

9:44

those automated workflows and things like that. So that's what we have been doing. Uh first uh we started doing that

9:50

with Dobby and I will talk more about it. We we shut down Dobby about a year ago and then we started a new company

9:57

Nimbus and we are just taking a different approach knowing uh a few

10:02

things and we'll get more into it but let me stop here like that's kind of the overall background and what was my

10:08

journey so far. Got it. Yeah, that's fascinating. So you do match it's kind of a matchmaking

10:13

idea. It is a matchmaking idea. So, Thumbtac and Angie's uh they are uh Cast Rabbit

10:19

they pioneered that matchmaking and they did Thumbtac um Angie in was founded in

10:25

1995. Thumbtac is also a 20 plus years company. So, every company is a product of that time. So, these companies uh did

10:33

what we call uh like you know they brought this industry from offline to internet but they make money once uh you

10:40

have that matchmaking right. So let's say Greg Michaelelsson is the homeowner and you submitted a job on Thumbtac and

10:46

or Task Rabbit. Your contact would be sold to five different providers and that's it. Thumbtac make money, right?

10:52

$50 per piece and they're done. They're not going to make sure the job is done

10:57

or like they're not going to get into the workflow itself and that's the difference. So we are not stopping there. We are first we're not going to

11:03

give you five pros like you know you can select one pro one job one pro idea and

11:08

we're not going to make money on like you know on that transaction on that matchmaking will be making money when

11:14

the job done with fivestar quality that's the big difference. Yeah. So how do you make money?

11:19

We uh in this new company Nimbus we make money as a subscription from service providers. So believe that if you can

11:25

power service providers with the proper architecture which is an autonomous system that is managing their enter

11:30

front office and enter back office. Front office is where they advertise their business and homeowners can book

11:35

rebbook refer them and back office is everything from that booking to quotation invoice communication payment

11:41

picture upload before and after and review management and everything. So that's you handle the payments as well.

11:47

Yeah payment absolutely payment is a big part of it. Got it. Well, that's fascinating.

11:54

Yeah. At lobby, we completed more than 10 millions in payment and like that is

11:59

uh like you know uh around 10,000 plus jobs and then on Nimba so far uh we

12:05

already crossed 100 jobs in uh in summer right after summer in fall and yeah so

12:10

we are very close to like you know 500 plus jobs. Uh what did I'd love to hear about kind

12:16

of what you learned from from Dobby to what is the new one Nimbus right? So uh Dobby was the first startup

12:22

and I um it was your first startup. It was my first startup and my co-founder at Dobby uh she worked at

12:29

Weiwork like you know she uh she was there for 5 years Alex Fondar. So she ran uh she was a director of operations

12:36

managing 14 buildings in New York. Uh and so she was coming from the operations side. I had the general

12:42

management background at data robot and data science background. And really this system engineering and the process thinking background that helped. Our

12:49

biggest learning was we became better first achieeper operators right and uh that was like you know kind of what we

12:55

took from Dobby and of course uh one of the core uh aspects of great founders is

13:00

that they know the industry inside out. I was scared at data at Dobby because uh I was moving from insurance and

13:07

completely entering uh entering a completely new industry and initially it was scary because uh not only a new

13:14

insurance industry insurance industry knew me so I have to give away cannibalize that like you know eminence in the industry and everything but I

13:20

understood that it's very portable the way we insurance industry also operationally extremely complex we

13:25

cannot just build a regression model we have to think about regulation and all of other things right that helped so uh

13:31

Dobby really The first thing is that better faster a cheaper founder and operator. Secing piece is a deep understanding of the home services

13:37

industry and our customers. Right? We were very fortunate to build a marketplace and it was uh you can still

13:43

go to app store and see Dobby. It is the highest rated home services app on the app store. So we got to meet thousands

13:48

of homeowners and thousands of providers and learned a lot from the users which is the most precious thing in any

13:55

industry. Uh third thing why Dobby did not work. There are two elements in it.

14:01

We were very profitable in our cohort of homeowners where they were like you know

14:07

in the low interest rate anyone who paid 3% or lower uh like you know mortgage rate because they had way higher

14:14

disposable income. What really happened like you know uh between 22 uh between

14:20

202122 and the years after is that the cost of home ownership went up significantly very quickly. And uh there

14:27

are four pillars of cost of home ownership. One is uh like mortgage which is non-negotiable and number two is

14:33

property tax non-negotiable property insurance kind of non-negotiable and utility those are also you cannot

14:39

negotiate with them. Third one is the renovation and maintenance and everything. So that bucket got squeezed

14:45

and we kind of started seeing although we are profitable within the cohort of low mortgage rate but as we try to grow

14:52

with these new homeowners uh they are not really spend sitting on a lot of money. We can go into detail but the

14:58

same home like the same home owners will be paying thousand to $2,000 more every month. That's a lot of money, right? So

15:04

that is kind of was the biggest driver of uh we could stay profitable in DC market with our cohort of homeowners

15:10

with a low mortgage but uh that was not the goal. We raised 6.2 million from incredible investors and our goal was to

15:16

really grow fast and build a venture scale business. We did not see a path forward of that

15:22

and then uh like you know Dobby was run by automation software but a lot changed in the like from like 126 software to

15:28

agent native agent native like you know software architecture so we wanted to move in there also. So those are the two reasons why we closed it and but but

15:35

with those learnings and so on. uh one of the things that I will say and and finish here like the one of the very

15:40

interesting learning we had from Dobby you should we should think about as a founder we should think about the

15:46

business model uh when things are going right not when things are going wrong.

15:51

So uh what are the things that we should uh thought like you know more of uh when

15:57

in 2122 we are growing fast in DC and getting those incredible reviews and see a profitable path really knowing that

16:05

what would happen if the interest rate goes very high very fast what would happen if inflation goes very high very

16:11

fast. Uh somehow like we should have done that and kind of we could have moved like you know did something

16:16

differently there. So really thinking about the business model even if everything is going well like you know really do that exercise every time like

16:24

you know once in a while like or pretty regularly that would be my like know one of the learnings. H

16:31

why'd you call it Dobby? Is it after the Harry Potter elf or Yeah, exactly. It is the Harry Potter elf. You're right. So we had our

16:37

vertical AI software that we built uh to run our entire automation inter operations was Gryffindor and our bot

16:43

was and like you know so yeah so we basically took different names from uh Harry Potter. Nimbus is also from Harry

16:50

Potter. Nimbus is the Yeah. The Nimbus 2000. Yeah. Nimbus 2000. Yeah. We just wanted

16:56

to choose a character that would not die this time or [laughter]

17:02

so sad. I remember reading that uh book six Dobby getting spoiler alert I guess

17:08

but just weeping weeping poor Dobby right but it Dobby was so like it was

17:15

heartbreaking uh for everyone but our customers our pros dobby was very smile generating name and everyone loved every

17:22

time and uh somehow we kind of this is our ICP ideal customer profile is the

17:27

Harry Potter generation so they immediately understood and it was fun that's fun uh how is the the whole rise

17:34

of uh LLMs and agentic stuff impacted the way that you guys are doing stuff at Nimbus.

17:39

Sure. Absolutely. So even it started impacting us at Dobby like I mean uh back in 2021 we understood we had enough

17:46

data to see different atoms and one thing we saw the most common question that a homeowner would ask us how much

17:54

something would cost right so home services industry is full of unknowns just like insurance industry people

18:00

don't know how much something cost even service providers do not know how much something cost you can still hear me

18:05

right yes I hear good yeah so the most common question was how much something cost that is 2021 we

18:12

discovered that so we added that feature on the in the platform that people would be asking us hey uh before I kind of

18:19

look for a provider can you tell me how much this would cost uh the first thing happened when in 22 September I think

18:26

chat GPT was published uh got released and by 2023 January we released this

18:32

thing called check the price so it was the first LLM implementation where you can ask any uh we already productize it

18:39

but Now we automated it thanks to JPT like you can ask the price and that great like [clears throat] it was so

18:45

popular like it was the most popular feature on the platform and we were not spending any money uh like before to

18:51

actually answer that question because when home owners are asking that question we are not making it it's not a revenue generating event right but uh

18:58

right so that's the idea so and and we built on that not only we'll be giving you the price but then we'll recommend

19:03

you these are the pros who can do it for you this was like you know the most popular feature we took like a kind of a

19:09

sidec car like a a giveaway to to get people exactly invested. Yeah. And a low commitment way

19:16

because anytime you write things like uh consumer is fascinating members all of these are consumer businesses. So what

19:23

you say to consumers matter a lot like am I when you like one of the reasons why we buy from Amazon because you know

19:29

return is free and return is no like no one is going to ask any question. It's a no question return. Same thing uh with

19:36

home services. If I click this book the pro, I want to know that I'm going to get a price. This would be fair.

19:42

Informed purchase decision matter a lot. So check the price was a huge level set because it was not just telling people

19:49

that how much something would cost but what needs to be done. So home owners do not know I'm I'm not that handy. You are

19:55

very handy. I remember in build carpentry and you are very you were a maker. But if you're not then like you

20:01

just don't know what needs to be done. So you'll be always thinking maybe someone is cheating like you know someone is not fair. But that it was

20:07

reducing those inform uh unknowns from the things. So that was the first consumerf facing features that we built

20:13

and we built on that and Nimbus also uh we just do it in a different way. Nimbus is way more advanced because it is the

20:19

agent native architecture and everything. So what do you mean by that? Uh it is like uh so when we built we

20:25

were Dobby was software heavy but uh but we build software for ourselves right.

20:31

So like you know our entire design of Gryffindor which would be running the operations for us we can see everything

20:37

the entire interaction between service and homeowners from the time of booking to all the way to five review we are

20:43

operating that software but and everything was architected according to that there would be a customer success

20:49

CSM customer success manager would be running this software right and the websites were built like human would be

20:55

reading it right like you know but now it's no more like we are designing everything that an agent will do it or

21:02

things will be running automatically. So that's in a very layman term like we can go in the deeper like you know how we

21:07

design that but that is in the in the highest level that is the idea. So if you look at we build a website for our providers they are not built on fancy

21:15

like you know kind of beautiful picture or anything they're very structured they're very kind of very easy for an

21:21

agent to come and understand that very easy for charged that and make it part of the AI search and everything. So

21:27

those are kind of decisions that uh we have taken ground up and that was a huge advantage to build ground up at Nimbus

21:33

and yeah and then uh one of the core use cases we are doing is automated generation of quotation and invoice. So

21:39

service pro spend a lot of time and they lose lot of customers for two reasons because they're delayed in sending the

21:45

quotation and number their quotation is incomplete but we are solving that problem with AI so it takes literally

21:51

second right now so when they're with customer they can send a quotation using our AI generated quotation and they can

21:57

edit it they do not want to lose control right this is a very interesting group of customers they want to have full

22:03

control so whenever we are implementing AI uh we are also giving them the full opportunity that you can review change

22:08

the price and change the like add one more labor, one more materials and do that. Yeah. So those are the use cases that we're doing. Everything else is

22:14

workflow automation, bud old RPA just powered by an ALM in the back end. Yep. Yeah. Yeah. Yeah. And the the

22:20

prices that you give people those are just like estimated like based on whatever is in the LLM.

22:27

That's right. It is based on LLM but uh insurance uh uh like know like insurance

22:32

quotation um home services also very location specific. So what something would cost in New York is different than

22:38

DC, different than Boston and different in San Francisco. So what we do we use different I will not get into exactly

22:45

what API we are using but we are uh running some Google search before like understanding some like getting some

22:51

fine-tuning data basically before we get into there we are still running it on the server we are not running it locally

22:57

but uh eventually once we have enough of quotation would be fine tuning with our own data that's the idea

23:03

got it now uh the the uh startup space for B2B software I imagine is is

23:11

different much different than the the B TOC type stuff that you're working on now. Um, Zerve, I don't know, Zerve is I

23:19

guess more of a B2D kind of play. So like we uh our users are data scientists, data analysts, people who

23:25

are interacting with data using code basically, right? And so we're, you know, we have like a free product and

23:30

people can sign up and they get certain number of credits every month. And uh, you know, one of the things that we

23:36

really value is providing something at at no cost to the user that's that's

23:41

valuable and useful and if they love it and they like it, then they can, you know, upgrade and get more credits and

23:46

and do more things and stuff like that. So that's kind of our our philosophy. Is this the same for you guys? Do you uh

23:52

how how is the BTOC space different from what we saw at data robot when we were doing the whole enterprise sales kind of

23:59

thing? Terrific question. So what you were describing is try and buy as opposed to we'll be showing you the demo and like

24:05

all that. The biggest difference what we saw in data data world was a classic B2B enterprise software and all that. So

24:12

what is the advantage of classic? Uh it's not easy. It's an advantage that of tailwind and headwinds like the

24:17

advantage is that we had a dedicated team that would be listening to our pitch. There is a dedicated budget that

24:24

would be like you know that we can access. Of course we have to navigate who is paying, who is using, who is protesting not to have deter and all

24:30

that. But there is a dedicated mind share like you know that want to listen to the new tools and new proposals. Uh

24:37

when we go to service providers or when you're trying to sell to homeowners, we don't operate like that. we don't have

24:43

today is my demo day like you know I'll be listening to new tools in my life so the only way to get in is try and buy

24:50

you just figure out and what the another way think about is that you cannot separate go to market uh GTM and and the

24:57

product anymore so you have to really embed your go to market inside the product somehow really so you can uh

25:04

like give the product to your uh target users and then they can use it and they

25:09

can see the value themselves and then they can buy how do we exactly do that Nimbus we scan through the enter

25:15

internet for we first identify who are the top local providers and let's say we found that Greg is one of the top age

25:22

pack person in Charlotte you don't live in Charlotte anymore but let's say Charlotte and we'll find that because

25:28

your data is publicly available from contact from task rabbit from ang and from other sources so we do that

25:33

systematically and then we accumulate all those review we aggregate all those reviews we build a beautifully

25:40

structured website where we make the booking rebbooking referring you the fastest in the industry and then uh we

25:47

give it to you you for free why because the marginal cost of creating one more website is zero to us like we have built

25:53

a 3D printer that we can do this for any providers anywhere in the country so that's a huge advantage of automation

25:59

right now historically you need three designers and like you know multiple engineers to do all of these things and

26:04

and and so on so that's a like that's how we do it try and buy and we can do that because of that 3D printer that we

26:11

have built side. Gotcha. So you give people a a free month or something like that to to give a free month. Yeah, exactly. And then

26:18

you see the value and then ask your homeowners are they find easy to book you and once you agree you are in like

26:24

$200 per month that is how we are starting it. We started uh with independent providers. So that's a $200

26:29

per month but once you go to the mid market and the larger players will be increasing that monthly subscription.

26:35

And how do you get people to see it? I don't imagine that many small businesses are like gee what I really need is uh uh

26:42

you know this pl like are there competitors in the space or or yes absolutely so so like I mean the

26:51

good news is that their phone numbers are available so real estate agents service providers these people want their phone number to be known so it's

26:58

not uh we don't have to run a lot of research to find their phone numbers but they're visible you can discover them

27:03

they're fragmented but they're discoverable so that's the first good news the Second good news is that people

27:09

feel a sense of pride when we have built a website for them. It looks good. It looks authentic and it is easy and we

27:16

can easily compare on thumb techch you need 20 different clicks and you are competing with five different pros. Three of them are low quality. Your

27:22

pricing and everything would be like you know kind of uh your profit margin would be squeezed here. It is yours

27:28

independent website that you own. You control take it and that's basically like you know kind of the idea. We do

27:34

have competitors and how we are totally different. Uh you don't don't need anything to generate a website in 2026.

27:41

We like generative is doing that. Our website is only 20% generating 80% like

27:48

you know it is essentially like you know kind of deep automation that is scanning through internet and millions of data

27:54

and everything and creating this very structured website and then every review we get we use generative AI to creating

28:00

an article. So that's the big difference. We see a common mistake that people who are not coming from in this industry, they are doing they're just

28:06

building a beautiful website, right? You know, but that website is not highly like converting because homeowners want

28:12

to hear from other homeowners. That means all that they care about is tell me how did you do something like this

28:18

and what other home owners are talking about. So our spotlight is nothing but reviews. It's not fancy like you know

28:24

landscaping has this beautiful landscaping picture. No, that's like generative AI would be doing that. We don't do that. We use generative AI but

28:30

it's uh generating articles and SEO optimized and AI optimized article for

28:36

every review we get we take that review take the job description take the before and after picture create an article automatically and things like that

28:42

that's what what kind of article what it's like so what you really service providers like you know there the there

28:49

are high quality pros and there are lowquality pros in the industry you do not want to work with lowquality pros

28:54

that's easy even with high quality pro what is difficult is that how do I know you are high quality pros. So every 100

29:01

job they do they're only collecting 20 reviews because you make writing review

29:07

is so difficult and you don't advertise them and review is your biggest marketing asset like you know so we

29:12

basically build our enter business on this philosophy that let's collect every single reviews and the same philosophy

29:18

is there in charity everyone wants to give money on charity but we make it difficult or for giving money to charity

29:23

but if you make it easy if you make charity at the point of sales that's why we see that in the grocery stores uh

29:28

then people will pay people will give money to charity. Same idea with review. Make it easy for home owners to write a

29:33

review, right? And then they write review and then we create that article in a way that it's not just SEO

29:39

optimized, it is actually AI optimized. So those articles are written as a first person. So chat give reference to first

29:45

person expert written articles. So we created that structure in the back end in a way that it is writing those

29:51

article optimizing for chat chip will pick this up and so on. First person is in the homeowner

29:56

the provider. The provider the provider. Okay. I got this job request from uh Greg Michaelelsson is a homeowner and I got

30:04

this uh request from him other day for water filtration system for his home and

30:09

oh my god this was so fun to work with Greg. Let me share a little bit more about this job and then before and after

30:14

picture how much you spend and what is the water filtration and blah blah blah and the entire thing is written on uh

30:20

AI. AI we are kind of keeping it as a fun. AI is choosing great incredible pictures that is just funny and

30:26

sometimes we kind of stop it but uh yeah but uh these articles are very valuable because it is giving them a lot of

30:32

exposure right got it and so are you just calling these providers up and saying hey this

30:38

is a service that's available or cold calling cold calling works like magic still uh cold calling text message

30:44

and and that's it like you know hey uh hey we built this website for you uh why because all of your reviews are you are

30:50

one of the top people but your reviews all all over the internet. We aggregated all of them. Now we have your own

30:56

website with all of your reviews from thumbtack, test, all everything in the same place. And yeah, do you like it?

31:02

What do you want to change? These are all editable websites and things like that. Yeah. And do they have access to change these

31:08

things or is it they have access to change this thing or they can send it to us? Uh so I will get

31:14

a little bit more into it. So our domain is hirenimbus.com, right? Why? Because

31:20

um you cannot sell technology to service providers. You have to sell outcome to service providers. So our pitch to them

31:26

just hire us to do your back end. You are the only workforce in this country which is doing both blue color and white

31:32

color. Why don't we take care of all the things back office all the thing front office and you do the thing that you do

31:39

the best. So you save time you save one hour every day. Either you spend that one hour with your family or you do one

31:44

or two more jobs in that time. That's it. That's high limbus to do your everything. Yeah. And what's adoption look like? Is

31:52

it is it going great? It is going well. It is going well. But this is still an artificial adoption because we have a head start because we

31:58

know the customers and we have a tremendous contact in service providers. So the true adoption because of Dobby.

32:04

Because of Dobby. Exactly. Because of Dobby, we know this industry so well and and all that. So nothing is free lunch

32:10

but we they there is a built-in trust with us and they know that we always help them and we we basically build like

32:16

service processor built their business ground up with Dobby so they know us very well. So that is an advantage. True

32:22

thing would be one of the biggest difference with Dobby and Nimbus uh after four or five years because we are

32:27

kind of a like old school marketplace and like that middle layer we were in DC

32:33

and San Francisco only two market and DC being the large market. uh Nimbus is in

32:38

uh 20 markets like we launched it in multiple markets at the less than six month and now we are slowly taking one

32:44

market and and going at what we call deep adoption like you know really visiting Austin visiting Atlanta and

32:50

like you know going to Charlotte and Raleigh and like that's what we're doing. So that's a huge difference. So the adoption is uh quite different in

32:57

like how we define adoption. So the true adoption is can we scale because our reason to stop Dobby is questioning the

33:03

business model and because it cannot really grow fast. So we have to prove to ourel we can grow fast with limbs based

33:08

in symbol model. So that's so far so good we are doing that that's why we launched m markets and so on. So that's been the adoption like you know it's

33:15

it's so far so good and right now it's still pre-product market fit if I'm being honest. Yeah. No makes sense. What's the hardest

33:22

problem to solve here or or what is talk to us about some of the hard problems? Maybe not the hardest but what's hard?

33:30

Right. So the hardest there are two really hard problem. The first problem is you asked this question a little ago

33:36

like do you what are your competitors? Is there any tool in this industry? This industry is actually full of tools point

33:41

solution. Uh people coming out of school I'm building AI reception is for service providers. Service providers do not want

33:48

to want AI to talk to their customers is scary for them right so that this industry is full of point solution. In

33:54

every batch of YC like you will see at least five point solution for home services industry. So service providers

34:00

are getting lots of calls. So that's like you know a very big challenge. So that's why this try and buy works like

34:06

magic and speaking their language works and so we basically have testimonials

34:11

from service providers who are working with us and you kind of you have to lead with those user generated content to get

34:17

their attention because getting their attention is difficult and they have some bad experience with thumbtack and

34:23

previous generation of marketing arbitrage uh lead generation company because they exploited service

34:28

providers. when you sell the same job to five different pros, four of them are not making money but spending money,

34:35

right? So they have bad experience with technology companies. So that's a big challenge. That's that's I would call

34:40

like that this first element that go to market and and getting their attention is the bigger challenge. The second one

34:46

is all like really deeply understand the workflow. Understanding the nuance of the home services workflow. uh what you

34:53

are very familiar with because of your banking experience, insurance experience that you have to be a subject matter

34:58

expert and it's not just CH problem it's not just hospital readmission problem that like you know that's a superficial

35:05

use cases but when you go inside an industry you have to deeply understand the users and the workflow so I would

35:12

consider that uh as a deep challenge and that is where also our mode because we

35:17

have been spending we spent like you know six years in this industry and we deeply understand and the users and and

35:23

the deep workflow. We run a home services company so we basically know what needs to happen to get a fivestar quality and things like that. So these

35:29

are the two challenges like you know uh what we have h fascinating

35:35

um what what's next what's the next thing on the road map for you guys what how so yeah so right now like uh really what

35:43

we are doing there are two elements in it and uh two core insight uh you cannot

35:48

sell a point solution to service providers and homeowners are also not going to buy a point solution right

35:54

what do you mean by a point solution so let's say we talked about check the price so what If I have a tool that

36:00

would just tell me what is the cost of a home service, right? You know, for a landscape, I need to replace my water heater.

36:05

I need to replace my water heater. So, you can build it and have some fun in the first few months. But then charge

36:11

GPT people will realize, oh, I don't need this. I can do this with charge GPT or cloud or anything else. So, um and

36:17

then on the service provider side, if you build point solution, they are duct taping right now five, six, seven

36:22

different solution. It is adding to their complexity, not really solving any problem. So the first thing what we have

36:28

been doing we are actually doing the entire workflow. You talked about payment not just payment like you know

36:33

booking rebooking and like know generating quotation whole communication uh home services is not like Uber you

36:40

don't just like click a button get a car you have to talk a lot with providers and home owners need to talk to each other and understand the problem and so

36:46

on. So how we have the inapp chat that just works you can upload video upload pictures and things like that and then

36:52

all of that. So we basically automated the entire workflow and there is a continuous work to be done there but uh

36:59

like you know that is the most important thing we did. However once we do the full workflow automation you have to

37:05

decouple them. What do I mean by that? Like let's say we we're running payment but um like if we force the service

37:13

providers that in order to do the payment you have to use Nimbus for booking you have to use Nimbus for

37:19

quotation then we are it is not a good idea. So people I just want to want to use Nimbus for payment. So we are now

37:26

decoupling all of these things and that's a product problem. So you have to still having a delightful experience and

37:31

capture all the data and everything. So that's the that's what we're doing and the really the next big step is that

37:37

both all of us are right now selling like Alex and I both are selling. So really sell this product to uh providers. Our goal is to have uh like

37:44

you know uh 200 customers by in the next 100 days. Uh so 200 paid customers. uh

37:50

we have uh around 50 customers right now. We just started like we went to text and finished in the mid December

37:57

and yeah so if we get uh 200 customers that is 40,000 uh MR. So the company is only a month old.

38:04

No no no company is uh no no we we completed text uh in December but we started in uh around summer. So June and

38:11

July is when we started. Yeah. Yeah. It's more than six months. Gotcha. And it's just you and your co-founder?

38:17

Uh no uh we have six people team. So uh we have two engineers uh three engineers

38:22

uh and then uh so Jeremy uh really helped me here. Jeremy helped me in many ways but this was a one big help. Uh so

38:30

Jeremy Ain our Jeremy founder of data robot founder of data robot Jeremy Ain. That's

38:35

right. Uh so Jeremy when I started doggy Jeremy got me connected with our Ukraine uh team over there and I uh built up my

38:42

engineering team in Ukraine when I started doggy and that that was a huge uh advantage and I continue with them

38:48

like we have now deep relationship I've been to Ukraine before co like you know six or seven times. So yeah, so we we

38:54

built Alex and I started Nimbus and then two of our operations team member joined from Dobby but the engineering team we

39:01

rebuilt and we have incredible team. Reinata does our design then Ian and Tamok there are running our engineering

39:06

and Fortune and then we have Bodhana, Alex and I like that is the team.

39:12

Awesome. And how about funding? Have you raised? Oh yeah. So we raised uh we did we're

39:17

keeping the PC undisclosed the amount but we raised like you know incredible like you know from incredible investors.

39:23

So our largest investor in the PC round is Chris Lanford who founded Lowe's ventures and so he started his fund and

39:30

like he invested and then progression fund were the Tik Tok uh early employees and early operators and they started a

39:38

company called musicly and then musically was acquired by Tik Tok. So Matt Lee uh is the lead investor there

39:43

and progression was our in third largest investors in Dobby. So they came back and then texters I started my founding

39:50

journey with texters and texters came back and invested in Nimbus and then Mark Artha was the first check at Dobby

39:56

our Mark Arita from data robot uh sales team and Mark uh is been Mark also the

40:02

first check at Nimbus. So like you know kind of Mark continued and yeah and all the Dobby investor will own 10% of

40:08

Nimbus. uh Nimbus is not taking anything from Dobby but uh without uh our uh

40:14

founding journey at Dobby there is no Nimbus like we learned so much and yeah so that's like you know kind of we call

40:19

it founder aligned equity. Mhm. Huh. Well that's amazing. It sounds like you're doing some amazing stuff. It's

40:26

hirenimus.com. Yes it's hirenimbus.com. Awesome. Awesome. How how about

40:32

personally? What's uh what you got coming up this year? You going on vacation anywhere? You you got any uh

40:37

exciting plans for the year? Sure, absolutely. So me and my wife Aperna we live in Washington DC. So

40:43

Washington DC met we live in Alexandria, Virginia, very close to Mount Bernal. Definitely make time to come here Greg.

40:48

It's beautiful and like uh we have been here for the last uh 11 years. Uh there was overlap also when Jeremy was here

40:55

during uh data public market and like you know kind of uh like you know when we had a lot of business with the

41:00

government back during COVID time. So at that time we were staying in the DC area. And now we're in the Virginia. Uh

41:06

just came back from a really good time off in uh in India and we also went to Thailand and Singapore in our way. So

41:13

yeah, so everything is uh going well. Personal life is uh going absolutely well. My wife just became a partner at

41:19

Betsswite uh economic consulting. She studied economics. She has a PhD in economics and she works at antitrust uh

41:25

practice over there. So yeah, so that's going well. Uh and uh she joined Bitswite I think uh 2015. So like it's a

41:33

like really good uh progress so far and uh she got to work on some amazing stuff like one of the biggest case was Amazon

41:39

Prime membership that was the highlight last year. So they were involved in that and

41:44

what what was the it was a case like a class action thing. Seemingly Amazon Prime uh kind of uh like made people

41:53

Amazon Prime members without giving enough information to the customers like you just you are just opting for faster

41:59

shipping without realizing that you just opted for monthly subscription now on and they had they called it internally

42:05

project Uklit because uh like project Iliad because uh it was so complicated

42:12

that the whole uh like you know like you know how company makes things

42:17

difficult for customer to cancel their subscription and they made it. Yeah. So they call it project Ilia because he be

42:23

reading a epic to like you know as a customer anyway. So department of

42:28

justice kind of put that case like you know hey you cannot do that and that is a very seinal case because a lot of

42:34

company kind of follow that kind of marketing technique. uh like uh yeah one of the people one of the like you know

42:40

companies I I respect a lot is HubSpot Brian Helik and Dha the founders Dha is

42:47

a big proponent of customer success and how how to make customers successful and he talks about it make cancelling things

42:53

easy like if you can get customers signed in in one click you should be you should be able to make like you know you

42:59

should have option to cancel the subscription in one click right so so that's a very big part be confident that

43:04

if people are like it's a very sustainable way to do business. So anyways, that's kind of how these things

43:09

work. I like that. It's kind of a don't be evil kind of thing. Yeah. Exactly. Yeah. And like this is a

43:15

big fun part of uh going back to work again like you know kind of consumer companies are so beautiful because of

43:20

that you kind of you talk to human in many ways enterprise also we talk to human these days like you know but uh

43:27

like you are not selling to corporations you're selling selling to users like you said you guys are selling to data scientists data engineers and like you

43:33

know but but really understanding uh like bringing that humanity in the selling process is very very important

43:40

in both ways. Well, uh I live out in rural Nevada, so

43:45

I don't think I can be uh hiring any uh Nimbus professionals if I need something done on my house. Uh but uh do you have

43:53

plans to go out to smaller markets? Is that something that can be done in an automated way? We have to learn more

43:59

about it like you know so what we have seen in the like a very good analogy here is that you have the all these

44:06

modern uh real estate companies brokerage or platforms like compass or broke open door and you'll see that

44:12

they're always staying with those major markets right so if I um like um if I

44:18

just have to give you an answer I don't see that we'll be at every corner of like you know kind of US uh uh like you

44:25

know in America but like if you really take a spectrum like you top real estate brokerage are in like you know 30

44:32

markets and then Uber is in a thousand markets we'll be somewhere in the middle right like you know kind of however

44:38

Nimbus is a model where we can easily geographically expand and like you know kind of uh even outside US so we can see

44:45

our doing because it's a kind of a platform play this is a problem in other continent and other countries Dobby was

44:51

very difficult to geographically expand but uh with that learning we kind of designed everything and with a

44:56

tremendous advantage autonomy and uh autonomous system and everything we can really expand everywhere. So really

45:02

don't know whether would be at the every rural part of uh the country or not but uh if we can build something that

45:10

homeowners find like not adding any cost to their life then definitely yes because it would be just making things

45:16

easier. Well Tandre thanks for taking the time I really enjoyed talking with you learning

45:21

about what you guys are doing. same like you know if it is okay I would love to know a little bit more about jerf like

45:27

you know can if you can please share a little bit more Greg yeah yeah so uh [clears throat] so

45:32

you're you're you've worked with data right I mean that was what we did at data robot this the problem really is that the the

45:39

stack for working with data is uh very challenging uh you know most

45:44

data scientists are working in tools like Jupiter or VS code or something like that and those are primarily like

45:50

local development environments like they live on your computer. Uh you know, they're not in the cloud. They're super

45:56

hard to collaborate with. Uh they're fragile. I mean, the first Jupyter notebooks were developed at Berkeley uh

46:02

to be used as like scratch pads in the classroom. And somehow that's become the

46:07

tool of choice for anybody that's interacting with with data using code. Uh because it's easy. You spin it up,

46:14

you know, it's interactive because, you know, the process of interacting with data is iterative. So it's not like you

46:19

just write a program and you're done. you have to, you know, draw a histogram and look at the distributions and see

46:26

the correlations between variables and what you learn at step one is going to inform what step two is and and so on.

46:32

So, you need that kind of interactive environment, but uh there hasn't ever

46:38

really been a tool that gives you that interactivity uh but also gives you the stability that

46:44

you'd need to actually build something that you can uh you can deploy. So engineers engineers [clears throat] are

46:51

constantly complaining to us about receiving these Jupyter notebooks. Uh you know like uh this data scientist

46:56

who's a who's a crappy coder you know wrote this notebook and now it's on my desk and I have to convert it into

47:02

something that can actually be deployed and used and and it's a big headache. So they're copy pasting code out of Jupiter

47:07

into VS Code and then you know figuring out the the CI/CD stuff and and how to

47:13

deploy it and everything. Uh what we've tried to do at Zerve is to build an environment where you can code in any

47:20

language you want, Python, R, SQL, whatever uh to interact with your data

47:26

uh but in a way that is super stable and collaborative uh and then we layered on top of that a very smart coding agent.

47:32

So uh with our agent you can basically just describe what you want and the agent will build it, run it and deploy

47:39

it for you automatically. Um so there are tools like this like lovable is is

47:45

really cool for building like the frontends of applications. Um you know

47:50

cursor is like and cloud code and some of these others are are good for like engineers uh you know um that are

47:57

building apps that sort of thing. But working with data is really different. So the zer agent can see your entire

48:04

workspace. It can see your data. It can see all your variable values. It can interpret the charts that it's drawn. uh

48:10

it can look at uh basically anything in your project and use that to build. So

48:15

like for example just earlier today I did a webinar where we uh we were optimizing the positioning of electric

48:21

vehicle charging stations uh in uh with some synthetic data

48:27

and we basically just one-shotted it. We said hey this is the problem. We've got you know a limited number of of charging

48:33

stations that we can put together. Here's the functional range. you know, simulate uh some demand data and then

48:40

run a linear optimization to figure out where the optimal placement of these

48:45

charging stations is. And within five minutes, we had the project completed.

48:52

Um so the agent agents have really kind of transformed the way that coding works and

48:57

it's it's remarkable. It really we're living in the future uh with the these large language models

49:02

and how they're impacting the way people do their work. and it is uh getting

49:07

better and better every day. So that is kind of the biggest part like you know whatever we are using today that would

49:13

be the like you know the the lowest amount of advancement we'll see in the years to come like you know kind of it

49:18

is just getting rapidly better. Uh so you guys also doing testing and like you know uh deployment and all of the other

49:24

stuff and maintenance that is that has been a big part of like you know production grade. Mhm. Yep. Absolutely. So uh you know you

49:32

can schedule jobs, you can deploy APIs, you can uh host streamllet apps, you can you know all those different kinds of

49:38

deployments uh those are those liveins and and it it just works or you can take your code and go home. So we we don't

49:44

have any kind of uh like lockin type features. We we also try not to be evil. So you can always dump all your code out

49:51

to a notebook or or a script and and just you know go somewhere else with it if you want to. But we think that having

49:59

used it, I would never code anywhere else. It's uh it's remarkable. It really is impressive uh what it can do.

50:05

And how did you meet the founders? Um well, so there's three of us. The company's actually based in Ireland. Um

50:12

the two other co-founders are are both Irish fellers, uh Jason and Philly. And

50:18

Jason uh Philly actually reached out to me on LinkedIn to pick a fight uh because he was seeing all the data robot

50:25

marketing stuff. Of course, this was after I left data robot, but he was seeing it and going, you know, I don't buy all this automated

50:30

machine learning stuff. Like, it doesn't it, you know, there's got to be more to it. Like, we don't we don't necessarily

50:36

believe this hype. Come come come and talk to us about it and help us understand what's going on. So, he really wanted to just kind of maybe not

50:42

pick a fight. I'm mostly joking, but he wanted to like have a conversation and understand things. Uh, and so, and I

50:48

ended up replying to him on LinkedIn. I said, "Look, you're right. you know, like there there is a lot of hype in the

50:53

whole automated machine learning space and it doesn't work for a lot of people

50:58

in and for a lot of cases. Uh, and so we got to talking and I I was an adviser for them for a little bit and then we

51:05

kind of stumbled upon the product that we needed to that we knew the market needed and we joined forces and and got

51:13

going. But Jason and Philly, what's that? It was a cold outreach. You did not know them. It was it was a completely cold uh

51:20

cold LinkedIn outreach. Uh I get I I don't know why I even responded to that one. I probably get 150 of them a week

51:28

on LinkedIn. And uh Philly's for some reason stood out. He's a he of course Philly is a sales guy. Like I mean that

51:35

was that's his background is sales. So you know I guess he has the the ninja level LinkedIn skills.

51:42

I reached out to my co-founder Alex Fonta like you know cold in London thing. There was no common connection. I

51:48

but I was a weiwork customer in New York and I loved and we all uh like know dream that you'll be spending time with

51:54

customers and if you work at wei work you'll be spending 100% of your time with your customers you'll be using the same bathroom restroom same coffee like

52:02

everything same hback so uh so I wanted to have someone uh like because it is such a customer like you know driven

52:08

company that the operations customer should be run by someone from weiwork so that was kind of and and she replied and

52:14

then uh we built dobby together and now we're building Nimbus together. Yeah. Cold outreach is legit like you know

52:19

what it can do. Amazing. That's wild. Yeah. Yeah. Amazing. And you are the chief product officer.

52:25

That's right. Yeah. Yeah. Unbelievable. And there is some like kind of philosophically it is very uh

52:31

kind of uh very aligned to the robot just more advanced and of course a product of time but uh like you know

52:37

philosophically it's like extremely kind of similar right. Well, it's I don't

52:42

know. I guess you could say in some sense it's aligned, but in in another sense it's completely in the opposite

52:48

direction because we we started out building for the experts and data robots message was always kind of like anybody

52:54

can be a data scientist and you know your grandma can build models that that sort of thing. Uh so we took a bit of a different

53:01

approach on that front. Um but we are focused on automation. uh we are you know utilizing these large

53:07

language models to make uh you know the the pipelining and and analytics and and

53:14

modeling work just work better and so yeah I I guess we've come at it from the

53:19

other end no no you're totally right because you have lovable versel client cursor replete so many tools and some of the

53:25

tools are talking to the software engineer like client is building for software engineers and that positioning

53:30

is very critical so totally totally agree with you because uh yeah Like we we we can go on and on because we know

53:37

that part of the data robot like you know when you send it for the citizen data scientist they think it's magic and

53:42

kind of we click a button get the outcome but the experts always take that approach that you were talking about is the iterative approach like uh like

53:49

there is no like only the useful model there is no perfect model so let's understand the data understand the data generating process and like keep

53:56

building iteratively and deployment and all that stuff amazing uh stuff thank you so much for uh inviting me Greg uh

54:03

this was incredible to catching up with you. I get your update through Gorav our common friend and

54:09

time. So thank you so much for like it was great talking to you and uh yeah so

54:14

we I was uh Razi u did uh this kind of a informal get together for dwood people

54:20

uh few I think in December for feature I I had to miss it. I was in Ireland when

54:26

it was it was so emotional. Oh my god. Like you know because it's like seeing people after many years but we all like

54:32

you know started at dat 2014 I uh and dro is one of a kind like you know it's kind of really showed us we also came

54:38

from large company travelers region bank and like so dot was true startup experience for all of us. So uh yeah

54:45

thank you once again and you have a wonderful year ahead. All right you too. Good seeing you.

54:50

Byebye. [music]

54:58

[music]