Full Transcript

·YouTLDR

Building AI Employees for Hospitality: How AITropos Takes Orders Where Customers Already Are

1:07:39EnglishTranscribed Jul 23, 2026
0:03

Welcome to Just Now Possible with Teresa

0:07

Taus.

0:09

>> Hi, my name is Santi Maruri. I am the

0:12

CEO at Itropos and I'm an engineer MBA.

0:16

I've been doing product software product

0:18

for the past 10 years and I'm an AI

0:21

fanatic. Absolute fanatic.

0:24

>> Love it.

0:27

>> I'm Juan Ao. I'm CTO at ITropos. I'm a

0:32

software developer. I've been doing

0:34

software development since the age of

0:36

around seven or eight. I'm 42 now. I got

0:39

my first computer when I was a little

0:40

kid and never stopped getting into it.

0:43

And with that in mind, I've always been

0:46

working with cutting edge technology,

0:47

always looking for the latest things. Of

0:49

course, AI excited me as soon as it

0:52

showed up. I have a degree in data

0:54

science and yeah I have a lot of

0:57

experience more than 15 years of

0:59

experience in hospitality software for

1:02

hotels and restaurants been working on

1:05

that. Yeah that's how we got where we

1:07

are now.

1:08

>> Amazing. One of my questions I want to

1:10

dig into is how why hospitality and how

1:12

you found this space. But before we get

1:14

there tell me a little bit about what

1:16

does atropose do?

1:19

Yes. So, we're actually building AI

1:22

employees for the hospitality industry

1:25

and we're trying to generate real

1:29

operational impact, right? So, it's not

1:32

just a bot, not just a chatbot, but it

1:35

has a lot of tools and we have many

1:38

integrations so that we can get done

1:40

real operational work. That's what we're

1:43

doing. We're addressing restaurants and

1:45

hotels mainly, but we also have bakeries

1:48

as customers. We're looking into bars

1:51

and every hospitality business can use

1:54

our AI employees basically.

1:59

Yeah, I love that. Okay. So, I we've had

2:02

actually quite a few companies on the

2:05

podcast where they're basically creating

2:08

AI employees in one way or another,

2:10

whether that's like customer service

2:12

agents or one of our recent episodes was

2:15

with a company that's like creating

2:17

agents to help with managing clinical

2:19

trials. I think this is a very hot space

2:23

right now because AI is capable of so

2:25

much, but it's also a little bit

2:27

>> of a tricky space, right? There's a lot

2:30

of fear around our jobs going away. Um,

2:33

and I think especially in the

2:34

hospitality industry, I could see there

2:36

being these are the types of businesses

2:38

where it's probably hard to find good

2:40

employees, especially if we're talking

2:42

about bakeries and restaurants, the

2:44

owners are probably tapped out and need

2:46

help. How do you think about this

2:48

balance of what's good for AI, what's to

2:51

do, what's good for humans to do. I

2:54

think especially in hospitality,

2:56

I think of hospitality and I im I

2:58

immediately think service. Like when I

3:00

show up to a hotel, I like having a

3:02

human welcome me.

3:03

>> Do you want to just do you want to just

3:05

tackle some of this? Like how do you

3:06

think of these hard challenges?

3:08

>> Yeah. Know and that's an amazing point.

3:10

And of course the AI employees are not

3:14

for every type of hospitality business,

3:17

but there are for there are super useful

3:19

for a lot of services. For example, we

3:21

are one one of our niches is QSR, quick

3:26

service restaurants, right? Because in

3:28

those cases, you don't go for the human

3:32

attention. You just go for food.

3:34

>> Yeah.

3:34

>> And the same thing applies for for

3:37

hotels. There's some hotels that of

3:40

course you have to have the human

3:42

attention, the human touch and there's

3:44

other hotels that you just want to get

3:47

some room service or that you just want

3:49

to schedule the taxi for to to the

3:52

airport. So that's where we believe

3:54

those are our targets. They are super

3:57

distinguished and the customer, the

4:00

restaurants and the business owners

4:01

really understand when these type of

4:04

services are useful for customers and

4:06

when they might uh need the human touch.

4:10

>> Yeah. If I'm sitting by the pool and I

4:11

need a margarita, I'm okay with ordering

4:13

from an AI. to your point of where it

4:17

impacts on places where this this

4:19

technology can be used. The places we're

4:21

looking to do it is not replacing where

4:25

a human should be but where the

4:27

technology is in the middle in between

4:29

humans. So basically instead of having a

4:32

human to human conversation through a

4:34

platform we know that we can replace

4:37

that interaction with human to agent or

4:39

human to AI through that same platform.

4:42

Right? So even more so some other places

4:44

where we think we can be very

4:46

competitive is where technology even

4:49

doesn't have a person on the other side.

4:51

So for example when you're using an app

4:53

we are basically are focusing on the

4:56

conversational interface uh rather than

4:59

the app interface.

5:01

>> Yeah.

5:02

>> Yeah. And maybe just to finish wrapping

5:04

up the idea our product of course we

5:08

market our product as AI employees. Erh,

5:12

we're even thinking about changing that

5:14

because AI is not our product. Our

5:16

product is delivering a high quality uh

5:20

service for guests and restaurant

5:23

customers and we actually do that of

5:25

course with a lot of AI but sometimes we

5:28

need humans to jump in, right? So our

5:30

main focus is to deliver an amazing

5:32

experience and one of our goals is to

5:36

pass the train test. The idea is to make

5:40

customers feel like they are speaking or

5:42

interacting with a human and we are

5:45

achieving that in many cases. Many cases

5:48

people the final consumers thank us and

5:51

even send us pictures food pictures to

5:54

thank us and they they think that they

5:57

explicitly think about how the attention

6:00

was super close.

6:01

>> You like this framing? It's interesting.

6:04

Let me back up. I can see clearly like

6:06

in hotels there's a lot of rules where I

6:10

could see your service being really

6:12

helpful. You already mentioned room

6:13

service. I joked about the margarita by

6:15

the pool, but anybody who's been to a

6:17

resort that's busy has had this

6:19

experience of you just can't even find a

6:21

human. Your example of a walk up

6:23

restaurant makes a ton of sense to me.

6:26

I'm curious about the restaurant

6:28

category in particular. You mentioned a

6:30

bakery. These are more experiences where

6:33

I feel like we're used to talking to a

6:35

human. And I know like here in the US

6:38

during COVID, a lot of our restaurants

6:40

move to QR code menus to reduce human

6:43

contact. But we're seeing at least where

6:46

I live, we're seeing all of that go

6:47

away. Like people want to connect with

6:50

humans. And but I also know restaurant

6:52

owners that can't find good employees.

6:55

they are strug they're like working two

6:58

full-time jobs and still struggling to

7:00

run the business. Tell me a little bit

7:03

about where do you see this playing a

7:04

role in restaurants just so I can get a

7:06

clearer picture of what types of roles

7:08

your product is filling.

7:11

>> Yeah. So let's jump right into a

7:14

specific example.

7:16

>> Yeah.

7:16

>> Let's say McDonald's.

7:19

you today you have to either order

7:22

through the kiosk or wait in line to be

7:25

served by a human.

7:27

>> Imagine if you were able to just get

7:29

into the McDonald's, take a seat and

7:32

order through a voice message on your

7:35

phone. And of course, the employee will

7:38

tell you when your food is done so that

7:40

you can pick it up at the counter or if

7:42

the restaurant can have a runners, they

7:45

will take you the food to your place to

7:47

your to the table where you're seated.

7:50

That's one of the main use cases that

7:52

we're aiming for.

7:55

>> Yeah, I could see this thing really

7:56

powerful. Like what im immediately came

7:58

to mind is my town has a big concert

8:01

venue in outdoor amphitheater and I

8:04

would love to be able to just order a

8:05

beer

8:06

>> there

8:07

>> and have it come to me rather than

8:09

walking to the tent and standing in

8:10

line. Yeah. Amazing. Okay.

8:13

>> Line there's a huge potential for AI

8:16

employees to taking care of their

8:18

orders.

8:19

>> Yeah, that's a great way to think about

8:21

it is why do we stand in line?

8:23

>> Correct.

8:25

All right, Juan, I want to go back to

8:27

something you said in your intro. You

8:30

said you've been in the hospitality

8:31

industry for 15 years. Is this how the

8:33

two of you landed in this space? Tell me

8:35

a little bit about how you found this as

8:37

the area you wanted to work in.

8:41

>> Yeah, actually we met with Santi, a

8:43

previous company we were working on that

8:45

was mostly related to marketing but did

8:47

marketing for hospitality basically for

8:50

hotels. Um but yeah, I got into that

8:54

company because of my experience also

8:56

working with hospitality systems. So

8:59

basically I work with the company here

9:00

in Argentina for around 15 years.

9:03

Actually it's more I always say 15

9:05

because I'm used to using telling that

9:07

but it's five years ago was 15. So we

9:10

could say that it was and I'm still

9:13

sometimes working with them like I do

9:14

consulting for them. So that that kind

9:17

of still there. But this company

9:19

basically builds one of the PMS

9:22

softwares, property management software

9:25

system, which is what hotels use for for

9:28

their operations. It's one of the one of

9:30

the system that has the biggest market

9:32

share in Argentina. And they also have

9:34

of course a POS system which is for

9:36

restaurants. So I've been working with

9:38

those systems for a lot of time for all

9:41

of these 15 years that I mentioned and

9:43

and that got me into kind of

9:45

understanding how the business or that

9:48

market works. So after meeting with

9:51

Santi and working on all of this of

9:53

course it was mostly

9:55

it went on its own decanting on that

10:00

specific market because of my

10:02

experience. Santi I know he can tell but

10:04

he's also been working a lot on that on

10:06

the industry as well. Previously he's

10:08

been working on another companies. So we

10:10

both know not had we both had the

10:12

knowledge uh maybe from different areas

10:14

but totally related both super nerdy

10:18

about the cutting edge technology super

10:21

nerdy about AI and so it was a

10:23

no-brainer to get started with this. Of

10:25

course, there were just because of the

10:28

contacts we had or what we've been

10:30

working on for these a lot of years. We

10:34

had a lot of easier entrance into the

10:38

market than if we had thought about I

10:40

don't know going into I don't know

10:42

automotive automobile market which I

10:45

know nothing about cars.

10:46

>> Yeah. So it sounds like you both had a

10:48

lot of domain expertise in this area.

10:51

you were geeking out on the technology

10:54

and just excited to play in this space.

10:56

One thing that's interesting to me is

10:59

you have a very broad problem space. So,

11:04

hotels have a lot of employees. They

11:06

have a lot of use cases. Restaurants

11:08

have a lot of employees, a lot of use

11:10

cases. Tell me a little bit about how

11:12

did you decide what to do first?

11:18

That's definitely a great question and

11:21

we did decide to work in the hospitality

11:24

industry but actually before that with

11:26

Juan we spent two years meeting with

11:29

industry experts and analyzing hundreds

11:33

of different ideas and we finally met

11:37

someone with a lot of domain expertise

11:40

in the restaurants area in the

11:41

restaurants industry and that's when we

11:45

realized Ed that there is one specific

11:50

use case, one specific feature that can

11:53

unlock a an immense potential which is

11:59

order taking. There's lots of companies,

12:02

lots of people helping with chat bots,

12:05

providing information etc etc. But

12:08

there's this use this specific use case

12:10

which is order taking orders which is

12:13

super hard and super valuable. So

12:17

actually we do have different employees

12:19

but our main feature is being able to

12:23

take orders and that's how we actually

12:26

decided about it. We actually started

12:28

doing a an assistant for waiters. That

12:31

was the first thing that we started

12:33

building. We literally spent six months

12:36

working on a device on a physical device

12:38

that would waiters would use and while

12:42

working in that we realized that taking

12:44

orders was the hardest part. So we said

12:46

okay let's focus on this and then when

12:48

we started offering this solution to

12:50

different restaurants to different

12:52

potential customers they started asking

12:55

for this service to be uh deployed in

12:58

for customers directly and that's how we

13:01

found out that there was huge potential

13:04

for a solution like this for customers

13:08

so we pivoted and we started focusing

13:10

specifically on that

13:11

>> and were you using AI already at that

13:14

time.

13:17

>> I can't remember a time when we didn't

13:20

use AI.

13:20

>> Yeah. Okay.

13:21

>> Anymore. I don't even want to think

13:23

about that. Gives me the chills.

13:27

>> Yeah. I think if you're familiar with

13:29

the product market fit question of how

13:32

disappointed would you be if this

13:33

product went away? I feel like AI for

13:36

people that have embraced it, like

13:38

disappointed is the wrong word. Like how

13:40

devastating would it be if this

13:42

technology went away? I think I wouldn't

13:44

be able to breathe.

13:46

>> Yeah, I know. Sometimes I wake up and

13:48

Anthropic has downtime and I'm like, how

13:50

do I do my job today?

13:52

>> Tony, we had a few episodes.

13:54

>> That happens to us. Well, we when we run

13:57

out of credits, for example, we use AI a

13:59

lot to code, as you might imagine, that

14:01

helps us move faster. And as soon as we

14:04

run out of credits, it's oh, what do I

14:06

do now? Do I have to code manually? No.

14:09

No.

14:10

>> Yeah. That's usually when I eat a meal.

14:12

I'm like, just step away from the

14:13

computer, go have a meal, go outside.

14:17

>> Yeah. Yeah. And it's fun because we

14:19

literally are the first ones to find out

14:22

when some of the LLM is not working. We

14:24

jump to X and there's nothing there.

14:26

Like silence, five minutes later, a

14:29

thousand tweets.

14:30

>> Yeah. Okay. So, you know what I really

14:34

like about your story is you clearly had

14:36

domain expertise. You still took a lot

14:38

of time to figure out the right problems

14:41

to solve. You found an area that was a

14:44

little bit com that that looked

14:45

promising. You started to build in that

14:47

space. Your customers, it sounds like,

14:49

pulled you even further and forget

14:52

waiters, do this for our customers.

14:54

>> Santi, you said you spent two years

14:57

looking for problems to solve. Was that

15:00

full-time? Were you both working

15:01

somewhere else? Tell me a little bit

15:03

about that exploration space. Yeah, and

15:06

two years is an understatement to be

15:08

quite honest. I spent the last 20 years

15:10

thinking about startup ideas, right? But

15:12

the two years was specifically related

15:15

to Huanu and myself, both of us

15:17

together. When as soon as we met,

15:20

>> we started enjoying very much our

15:23

conversations and started thinking about

15:25

so many ideas that we could build.

15:27

Actually just a very quick uh we we love

15:31

both of us love astronomy and at some

15:33

point we fantasize with building a

15:36

company that is called TAS which was

15:39

telescopes as a service.

15:42

>> Nice.

15:42

>> We were trying our idea was to use

15:45

SpaceX to h put in orbit a telescope and

15:49

lease the time of the telescope. So

15:51

that's how we literally spent a lot of

15:54

time together thinking about ideas. We

15:57

were working full-time. So, yeah, we did

15:59

it on our spare time.

16:01

>> Yeah. I hope you someday also make that

16:04

telescope company. I think that would be

16:06

fun.

16:07

>> Yeah, absolutely.

16:09

>> Okay, so let's get into this a little

16:11

bit. You got pulled into you were first

16:13

making software for the waiter to make

16:15

it easier to take orders. This got

16:17

pushed into can we just give it to the

16:20

customer. The first thing I love is that

16:24

this is a very specific use case. You're

16:26

not looking at a hotel and saying,

16:27

"Let's do all their jobs for them."

16:30

You're saying, "Let's take orders." The

16:32

other thing I love about this is Santi,

16:34

you mentioned this was a hard problem

16:36

and I can imagine it's not you're not

16:38

just spinning up a knowledge base and

16:40

being an answer bot. You've got to

16:42

integrate with point of service system,

16:44

point of sales systems. I'm imagining

16:46

you're interfacing somehow with a

16:48

kitchen that maybe is making food or

16:50

something real in the physical world

16:52

where that order is turning into

16:54

something real. So give me a sense of

16:56

what does it take to solve a problem

16:58

like this? What's the big picture?

17:02

>> Yeah. So the connections, the

17:05

integrations, all of that, although they

17:07

are quite complicated, those are not the

17:10

hardest part.

17:12

The hardest part is to being able to

17:14

translate the not deterministic world of

17:18

human conversations and LLM's into a

17:23

structured information so that you can

17:26

feed that to systems. That is one of the

17:29

hardest parts, right? So that's what

17:31

took us a long time. It's you can do a

17:34

prototype in a day. Anybody that uses AI

17:37

can do a prototype in a in a day, but

17:40

making sure that this is consistent and

17:42

that works every time takes a lot of

17:45

time. That's a one of the things that we

17:48

we realized when we started working. And

17:50

how did we solve that?

17:53

Putting a lot of hours. putting a lot of

17:55

hours understanding and making and

17:57

making a an architecture that is super

18:00

advanced. I'm gonna let Huanu speak more

18:04

about all the architecture that these

18:06

agents are using. They are not a simple

18:08

h a simple prototype that you can build

18:11

in a day.

18:12

>> Let me make sure I understand where you

18:14

said the challenge was. So there's first

18:17

you mentioned the non-deterministic

18:18

human, which I love this because we all

18:20

talk about the non-deterministic LLM,

18:23

but it turns out humans are also

18:25

non-deterministic. Yeah.

18:26

>> And there's this like first layer of I'm

18:28

assuming this is a chat interface. I can

18:31

say anything.

18:32

>> Yeah.

18:33

>> Yeah. You connect it to WhatsApp.

18:36

>> Okay. So that's the first challenge of

18:38

the user can enter anything.

18:41

>> Yeah.

18:41

>> Yeah. The platform is actually channel

18:44

agnostic. So basically we started with

18:47

WhatsApp here in Latin America is like

18:49

the OS the operative system of Latin

18:52

America that's what we use but we could

18:54

easily connect to iMessage SMS any other

18:57

channels and it's part of our road map

18:59

as well and it's the easiest thing to

19:01

connect for us. We're just using

19:02

WhatsApp right now because of that. But

19:04

one thing I wanted to mention regarding

19:06

what Santi said about the challenges

19:09

was all of the things that Santi said,

19:12

not only they had to be correctly done,

19:15

but as you might imagine, since you're

19:17

doing realtime order taking, the agent

19:21

has to be fast and responsive and

19:24

respond correctly while in a time

19:27

fashion where it doesn't get the

19:29

customer waiting. As you might imagine,

19:31

agents as agents today are mostly used

19:34

for longunning tasks, right? And that

19:37

takes a lot of time to process

19:39

information. So, one of the biggest

19:41

challenges and I had Salty like here on

19:43

my ear constantly. We need to look I

19:46

remember I I don't know if you saw the

19:48

playlist series based on the sp how

19:50

Spotify was built. So,

19:53

>> okay. So the playlist is a series where

19:56

it shows how Spotify was built from the

19:57

different perspective of the builders.

19:59

And one of the things you could see is

20:01

how the person had the idea, I can't

20:03

remember his name, constantly had the

20:05

technical guy saying faster. I need

20:08

songs to load faster. No, the time

20:09

between a song and another song has to

20:11

be faster, faster. That's how I felt

20:13

with Santi on my side. But that paid off

20:15

because we're actually right now at a

20:17

place where like Santi said, people are

20:19

not noticing uh they're chatting with an

20:22

agent. Not only because of the response,

20:24

the way that the agent responds, but

20:25

also because it responds not too fast,

20:28

but not too slow either. And that's kind

20:31

of part of the challenges that we were

20:34

trying to solve. I love that you

20:36

mentioned not too fast as part of the

20:39

challenge because I know like when I

20:41

send a support email and I get a really

20:43

detailed response one second later I'm

20:45

like yeah that was an AI.

20:48

Okay. So it seems like there's this

20:50

first layer of the human gets to enter

20:52

whatever they want. So you got to deal

20:53

with the messy of the messiness of this.

20:56

You have an agent that is trying to

20:58

understand that message. I imagine your

21:01

agent is doing the heavy lifting of

21:03

interacting, integrating, like

21:05

structuring that input in a way that

21:08

works with your now deterministic

21:10

systems, point of sale, whatever. So,

21:13

give me a sense of I want to go back to

21:15

like day one. What was your first

21:18

prototype? How did this start? How did

21:19

you even evaluate if AI could do any of

21:22

this?

21:24

Yeah, I that's you can start but one

21:27

thing I want to say is that the core

21:29

piece was always the same one which was

21:32

basically had an integration with this

21:35

external system that was something that

21:38

remained along these different

21:39

iterations but I let Santi talk about

21:42

the first product we built that we

21:44

iterated actually we had uh two I think

21:47

we're on the third iteration right now

21:49

right Santi we have first the hardware

21:52

then we have the custom app for waiters.

21:54

And now we're actually at the uh the

21:57

end.

21:58

>> Yep. Yeah. I can't remember how many

22:00

iterations we've done on this product to

22:04

be honest.

22:05

But yes, so we always we were super

22:10

optimistic about this, Teresa. We were

22:12

super optimistic. We saw the potential

22:15

of AI and we actually never thought that

22:19

this couldn't be done. I think it was

22:22

our our determination to make this work

22:26

which made us push very hard and again

22:30

the first time you one of the best

22:33

things about AI is that it gives you

22:35

some very quick dopamine hits because

22:38

making a prototype it's awesome. It's

22:40

awesome. But that is good and bad at the

22:43

same time because you have no freaking

22:45

idea what you are what you're starting

22:48

to do, what you are getting into. You

22:51

have no idea before you start, but it

22:53

gives you this dopamine heat and it's

22:55

like you feel a superhum and you're

22:57

convinced that you can do it. That was

22:59

where we were. But I honestly had some

23:03

doubts along the progress.

23:06

We spent a few months working on it and

23:09

we still had a an unacceptable error

23:13

rate because we wanted to make this

23:17

perfect and that's when we started yeah

23:21

testing different ideas playing around

23:24

with so many tools so many different

23:27

agentic architectures

23:30

we have five different types of ag rack

23:34

with different It's it got complicated

23:37

at some point. We started thinking about

23:39

the physics and how it should evolve etc

23:43

etc but yeah it was super super hard to

23:47

to be able to really to understand every

23:50

time what the customer is ordering

23:53

especially in especially in different

23:55

restaurants that probably have products

23:58

that are quite similar. So if you fit

24:00

that to a prototype you're done. That's

24:02

when you say, "Okay, this is might be

24:05

harder than it looks." But yeah, but it

24:08

was a hard process, but we're right now

24:12

it's working so good that we are very

24:15

proud of what we built.

24:17

>> Yeah. One one thing to mention is, and

24:19

this is anecdotal, but like Santi said,

24:22

sometimes you could be like overwhelmed

24:24

at things not working and like you start

24:26

having questions. like something said

24:28

like we we never thought that this

24:31

couldn't be done. We just were thinking

24:34

whether we were at the right time and I

24:35

had this memory I have this like

24:38

snapshot of a chat I was having with

24:40

Santi chatting with him where we were

24:42

doing this uh second iteration. So the

24:46

first iteration was a hardware that was

24:49

the idea was that kind of like having a

24:51

headset for waiters where they would

24:53

just talk to an agent, the agent would

24:55

help them etc. That was super hard not

24:57

only because of the model but also

24:59

mostly because of the hardware build.

25:01

Second one was something similar but on

25:04

a on an app where it was mostly like a

25:06

chat app and then the waiter would just

25:09

make the order take the order from the

25:10

customer make the order on the app and

25:13

then generate the order and send it to

25:15

the POS. Like I said all iterations had

25:17

the core idea of integrating with the

25:19

system. Now this third one is with the

25:21

chat. So when we were at the second

25:23

iteration, we were trying to get our

25:27

agent or agents to build an order right

25:31

with uh orders are super complex objects

25:34

in data terms because you have the

25:37

product the product can have a variation

25:40

in the recipe. The product can have a

25:42

something called a modifier which is

25:43

like large small. The product can have

25:47

extra products linked to it. So you

25:49

could have like a promotion. If you buy

25:50

two products separately, they have a one

25:52

price, but if you buy them together,

25:54

they have another price. And depending

25:56

on which POS you have, all POS have uh

26:00

different data structure. And that's

26:02

actually a regular problem of POS

26:04

system. Like

26:06

I tend to think that POSOS systems are

26:08

still an unsolved problem because they

26:10

all have different ways. Each restaurant

26:13

or each venue has a lot of each of their

26:16

own different ways of doing things each.

26:20

So they all have in the end to ask for

26:22

specific custom implementations to the

26:25

so to the software that the company that

26:27

developed the software. So there's a lot

26:29

of as you might imagine there's a lot of

26:31

variance and working on implementation

26:34

on integration with all of them for us

26:36

it's super hard but that wasn't the

26:39

hardest part. So when we were doing this

26:41

second iteration and we couldn't get the

26:44

agent to correctly build this system, I

26:46

was like Santi remember just a message

26:48

from Santi Hanu I'm not sure can we do

26:52

this it's not fully working as expected

26:55

and then I was like trust Santi we're

26:57

going to make it uh and in the end it

27:00

turns out that it's about not giving up

27:02

because the only limitation is the

27:04

technology I remember having like I was

27:06

saying is most of the times is basically

27:08

trusting your product and understanding

27:11

whether the problem is that can it be

27:13

done or not and if you think it can be

27:16

done what's keeping you from it and it's

27:17

most of the times is if it's the

27:19

technology that's limiting you is

27:21

whether do you think the technology will

27:24

be there at some point or is it like a

27:26

physical limitation I think even Ilan

27:28

Musk works with this idea or whether if

27:31

it's not allowed by physics then it's

27:34

can't it can't be done but if physics

27:36

allow it that you can do it we're super

27:38

far away from that. But I was I remember

27:41

having Santi sending me this message

27:43

saying Juano I don't know if we can do

27:45

this you it's not working as expected

27:47

we're taking a lot of time and I was

27:49

like let me see Santi trust trust we can

27:52

do it and just that exact same date one

27:57

of these companies that we use for model

27:59

released a news model smarter and it was

28:03

just a matter of let me try with this

28:05

and I just switched the model and it

28:08

started working like without even

28:10

changing anything from our side. Of

28:12

course, there were a lot of nuances that

28:13

we then fixed and improved and we

28:15

constantly keep doing that. But it was

28:17

just a matter of waiting for the

28:19

improvement on the technology, the base

28:21

technology that we were using to have

28:23

our product fully up and running. And

28:25

that's when we said, "Yes, this can be

28:27

done." And the fact that we actually got

28:30

to the point where the model that we

28:32

needed showed up for us to get moving

28:36

faster or actually get moving made it,

28:38

okay, man, we're on the right track and

28:40

we are early.

28:41

>> Yeah, this is I love this because I

28:43

think I think it was Andrew Karpathy

28:45

said it might it was either Andrew

28:46

Karpathy or it might have been Boris

28:48

Churnney from Anthropic said to build

28:51

your product for the model that's coming

28:53

out six months from now. And what's fun

28:57

about this space is that like you build

28:59

something and then time passes and even

29:02

if you do nothing, your product gets

29:04

better. Like the brain behind your

29:06

product gets better. And it's so cool to

29:09

see what this unlocks. Like I had this

29:11

experience myself with the product that

29:13

I'm building. I was like prototyping in

29:16

clawed code and Opus 4.6 six was doing

29:19

the heavy lifting and Opus was great but

29:21

Opus is very expensive and you cannot

29:23

put that in a production product and so

29:26

I like switched to Sonnet I started to

29:28

play with Haiku it was okay and I was

29:32

like that's okay time will pass we will

29:35

get opus level brain at ha coup prices

29:39

keep going and I feel like this is a

29:42

brand new part of building products like

29:44

I can't really think of anything else

29:46

where this has been true where like you

29:50

really can start to see the future and

29:52

build right on the edge of that future

29:54

which is really fun.

29:57

>> Yeah, for sure. The only caveat I would

29:59

say is that you got to work Yes, you got

30:02

to work on the thing that is coming. But

30:06

you have to make sure that you build it

30:07

correctly because if you only expect

30:09

your product to get better just because

30:11

of the model, that's fine. But that's

30:15

just adding steroids to something that

30:17

if your product is not as performant or

30:20

as good, maybe in the end it will fail

30:23

because maybe not because it's not

30:25

working but maybe because of costs as

30:27

you just said like intelligent smart

30:29

models are expensive. So if you only

30:32

wait for super smart models to come out

30:35

and you just depend on that it's going

30:37

to super expensive. You still have to

30:38

make a lot of engineering and

30:41

architectural decisions within your

30:42

product to make sure that your product

30:45

should and it does the right thing

30:47

internally to avoid these other problems

30:49

that might have.

30:51

>> Absolutely. And of course there's the

30:52

hard challenge of how do you know what

30:54

the model's going to get good at and

30:55

what to delay for later versus what to

30:58

work on for now. Okay, let's get a

31:00

little bit into give me the high level.

31:03

You've mentioned agents a few times.

31:06

walk me through. I'm a customer. I'm

31:08

trying to place an order. What happens

31:10

next?

31:16

>> So, basically, you just open up. In our

31:18

case, we're just going to talk with what

31:20

we currently have. You just go to your

31:22

WhatsApp, which is basically something

31:23

you have installed. That's one of the

31:26

great things about this approach that

31:28

we're taking is that no one has to

31:30

install anything at all. They just have

31:32

to use whatever they use. Uh, and then

31:34

you basically go to the contact that

31:36

represents the restaurant and you start,

31:38

hey, I want to make an order and the

31:40

agent starts talking to you and also

31:42

guides you on the way on what you want.

31:45

It even does recommendations for you. It

31:48

if you have some sort of idea of what

31:50

you want to eat, it recommends based on

31:51

your ideas. It tries to also match

31:54

products that you like with other things

31:55

that could go well with that. Of course,

31:57

there's a lot of marketing from the

31:59

venue side that they, hey, I want you to

32:01

offer this when they ask for this other

32:03

stuff. All of that kind of things that a

32:06

restaurant or a venue would actually

32:08

train an employee to do. Our agent can

32:10

do it also. The good thing is that the

32:12

agent doesn't get tired, doesn't get

32:15

stressed, they're never depressed, etc.

32:17

All those benefits. But from there

32:19

basically the or the as the user and the

32:22

agent chats basically the agent the

32:24

steps that it takes it starts building

32:26

an order with the products that are

32:30

being add that are being requested sets

32:32

like we mentioned the modifiers comments

32:35

variations on the recipes

32:38

there's a lot of internal processing on

32:41

because we have many features internally

32:44

while that happens which are like when

32:46

you add a product make sure that it has

32:48

stopped stock. Uh when you add the

32:50

product, make sure that it has stock,

32:51

but also you need to check whether the

32:53

order is for today or is programming

32:55

scheduling for another date. So if

32:57

you're scheduling for another day, you

32:59

have to make sure that stock it will be

33:01

available for that day. And so there's a

33:03

lot of these rules that I like I was

33:05

mentioning have to happen at the same

33:06

time in parallel most of the times to

33:09

make sure that we can reply in a

33:11

fashion. Once you do that, you build the

33:13

order. The customer is happy with their

33:15

order. Then the agent basically asks

33:18

whether it's for takeaway or for

33:20

delivery. And it's for takeaway, they

33:22

just ask for the name. If it's for

33:25

delivery, they ask for an address. And

33:26

we check the availability region for

33:29

delivery. So the agent can also know

33:31

whether they can deliver to that address

33:33

or not. And this cannot happen actually

33:35

in the middle of the conversation. It's

33:37

not something switched to when the order

33:38

is finalized. It's like when you talk to

33:41

a person, you just, hey, do you guys do

33:43

deliver to this area? Yes, we do. No, we

33:45

don't. So maybe even if the order is not

33:47

set, it can answer that question. Uh

33:49

sometimes even our agent is used for

33:51

asking questions about venue, not really

33:53

ordering, which also works. So that's

33:55

standard use for agents. And so once you

33:58

have the order and you have the delivery

34:00

method whether it's pickup, takeaway or

34:03

delivery, the agent basically closes the

34:06

order and generates a payment link which

34:08

is also sent through WhatsApp. And this

34:11

payment link basically the person just

34:14

clicks on it, goes to whatever payment

34:16

platform the customer use, they make the

34:18

payment and then on WhatsApp they get a

34:21

message, hey we receive your payment,

34:22

your order is being processed. So after

34:26

that the c the venue start preparing the

34:28

order which also triggers a notification

34:30

to the customer on their messaging

34:33

saying hey your order is being start

34:35

started working it will be around ready

34:37

in around 25 to 50 minutes whatever that

34:40

time is and if it's for delivery you

34:42

will tell them hey your order is ready

34:43

and it's out for delivery to your place

34:47

or your order is ready you can come and

34:48

pick it up and so basically as you can

34:51

see this is what currently delivery apps

34:54

are doing, but in a way that the person

34:57

doesn't have to leave their their happy

35:01

their safe place, right? Which is their

35:03

messaging app. They don't have to go to

35:06

external apps. Here in Argentina, we

35:09

have these two apps that are for order

35:11

delivering. I know in the US, you have

35:14

Door Dash and even some some venues have

35:17

their own custom app for order ordering,

35:20

which is a mess. As you may imagine, you

35:22

have to install external apps just for a

35:24

single venue. So, what we're trying to

35:26

do is like Santi said is just improve

35:28

the experience of the person when taking

35:30

orders by not leaving their comfort zone

35:34

so to speak for on their phone. And

35:37

that's basically it. We have the I think

35:40

the only time they leave the app is when

35:42

they actually just click the payment

35:44

link which takes them to whatever

35:46

payment platform it is and then that's

35:48

it. the remaining of the experience is

35:51

fully integrated into their chat

35:53

experience. It even tells you when your

35:55

delivery guy is at your door.

35:58

>> Help me understand this in the context

36:00

of let's say I'm ordering like food for

36:02

takeaway.

36:04

Am I going to the restaurant website to

36:06

see the menu, but then I'm ordering

36:08

through WhatsApp? Like I can imagine if

36:10

I'm at a hotel, I have a room service

36:12

menu in my room. I'm looking at that.

36:14

I'm using WhatsApp to order. But is that

36:16

the idea? like instead of I'm still

36:19

looking at the restaurant menu online

36:21

and then using WhatsApp to order.

36:25

>> That depends a lot on the user. The user

36:27

can either go to the website to see

36:29

previously what there is for to then go

36:31

and order or they can ask the agent to

36:35

either recommend them, tell them what's

36:37

in stock, or the agent can even send. So

36:39

our agent actually can send attachments.

36:42

So they can send you a PDF with the

36:44

menu. They can send you pictures of the

36:46

food of course as long as they are set

36:48

up on the system. But so the full

36:50

experience can h experience can happen

36:52

on the app

36:54

>> or they could just go to the website and

36:56

see the menu. That's up to the customer.

36:57

We don't force anything on that end

36:59

>> where the customer gets the what the

37:03

number to speak with our agent depends

37:05

very much on the business. There are

37:08

some businesses you just mentioned that

37:10

you have you can have the QR code on the

37:12

on on the room in a hotel

37:14

>> or you can have the QR code printed out

37:16

on a table in McDonald's for example or

37:19

you can get the number from the website

37:21

or even you if you're a recurring

37:24

customer you have it you already have

37:26

the connection so you just have to go to

37:28

whatever messaging platform you're using

37:30

and and search the name of that

37:32

restaurant. I can imagine too for like

37:36

delivery, people have their go-to spots.

37:38

Like I know exactly what I want to order

37:41

from a specific restaurant, so I don't

37:42

need to look at a menu. And I can see

37:44

that being a really powerful use case as

37:46

well.

37:46

>> Correct.

37:47

>> And I love that it's through WhatsApp

37:48

because I don't

37:50

>> I definitely don't want to call a

37:52

restaurant ever. I'm not a millennial,

37:53

but I feel like I have that millennial

37:55

trait. I just don't want to call

37:57

somebody. And I also I want to see I

38:01

want feedback that my order is correct.

38:03

I hate placing an order on the phone. I

38:05

don't really trust that they got my

38:06

order correct. And I really think the

38:09

world should just operate over text. So

38:11

I think this is amazing.

38:13

>> It's basically text, right? We just have

38:14

a technology that basically converts the

38:16

audio into text. But the idea is giving

38:19

the full conversational experience of

38:22

DMing through WhatsApp you have with

38:24

your friends. So whether you want to

38:26

chat or just send an audio message,

38:28

which is right.

38:29

>> Yeah. Okay. So

38:31

>> you you would clearly be one a good

38:32

customer, a customer that uses this

38:34

tool.

38:35

>> I would be I wish all things could be

38:38

ordered via text. Okay. Let's get under

38:40

the hood a little bit. It sounds like

38:41

there's a lot you're orchestrating

38:44

behind the scenes, whether it's checking

38:45

delivery zones, checking stock, making

38:48

sure you have all the right data for the

38:50

point of sale system. How does this what

38:53

does this look like under the hood?

38:55

>> Good question. I don't know if you want

38:56

to tell your secrets. No, just kidding.

38:59

No. Okay. These are different. So,

39:01

basically what's going on under the hood

39:03

is just to give a quick sample is

39:05

basically we have our system just

39:08

receives a web hook or a call from

39:10

whatever API we use to message for the

39:13

messaging channels. In this case,

39:15

WhatsApp receives a message and from

39:18

there starts a full pipeline that does a

39:20

lot of things.

39:22

The main challenge was so we actually

39:25

the first thing we had was that the

39:27

pipeline was straightforward like one

39:28

thing after the other right just to make

39:30

sure that things worked but on upon

39:33

iterations on that and we started saying

39:35

okay we need to shrink down the times I

39:38

think the biggest challenge was actually

39:40

figuring out which parts for example

39:42

could be parallelized right so

39:44

>> you got do we talk directly with the

39:47

POSOS as we build the order or do we

39:49

build an external system that takes the

39:51

order, which is much faster. Of course,

39:53

we the first time we just went through

39:55

the integration part. Next iteration was

39:57

a no-brainer. Let's just do everything

39:59

inside our app and then send to POSOS or

40:02

the integration. The next one was how

40:05

many things can we do in parallel that

40:07

can be done in parallel that doesn't

40:09

need a sequential sequential processing.

40:12

Right? So for example, if you're

40:13

ordering for if you're searching or

40:16

ordering for multiple products, the

40:18

agent can basically search for all those

40:20

products at the same time and then build

40:23

a response based on all the results,

40:25

right? So instead of searching one by

40:27

one, you do a lot of multiple searches

40:29

at the same time. Then the other thing

40:32

is like database the database the

40:34

basically the database infrastructure

40:36

how powerful is the database or the

40:39

database choice that you use so that it

40:41

actually has quick results and ordering.

40:44

Uh but all of that is basically

40:46

architecturing different ways of

40:49

treating the data and parallelization

40:52

caching database infrastructure database

40:56

engine of course yeah for different

40:58

things. So all of those things had to be

41:00

considered at the same time which is I

41:02

think the hardest part but in the end

41:04

what happens is that the agent in our

41:07

scenario we're using tools and we

41:09

decided to use agent tools rather than

41:12

MCP or rag because it's the fastest most

41:18

efficient way for the agent to interact

41:19

with MCP it would have to basically go

41:22

through the MCP understand what's going

41:23

on make a request to an endpoint to

41:26

which could potentially be fast But

41:28

still it's one extra step with tools.

41:31

It's basically okay call this function

41:32

and the tools are already on the prompt

41:34

of the agent. Now we do some rag some

41:38

initial rag retrie augmented generation

41:41

for knowledge bases or if we want to

41:44

preload information sometimes.

41:47

So we have these hacks that we've been

41:49

implementing where for example our

41:52

system prompt is built based on the last

41:55

message and the previous messages. So we

41:58

got two system prompts. One is like the

42:00

main system prompt and then we have

42:01

something called like updated system

42:03

prompt which is like an immediate system

42:05

prompt that the agent gets to know more

42:09

information about what's going on but

42:11

from the system role and that basically

42:15

builds the prompt based on okay so

42:18

you're looking for this product so let's

42:20

quickly build into that small system

42:22

prompt information about that product

42:24

before the agent can respond so that it

42:26

doesn't have to figure out a tool to use

42:28

and search for it. We just fitted that

42:31

information right away because we have

42:32

to figure that out. Right? So all of

42:35

this all of the like I was saying it's

42:37

mostly architecturing engineering but

42:40

just coming up with good ideas on how

42:42

can you resolve these problems. Hey,

42:44

it's taking too much time on a database

42:46

call. Okay, can we do this faster?

42:49

Figure it out on our own

42:50

programmatically instead of having the

42:52

agent figure it out and figure out which

42:54

tool to call. So

42:56

>> yeah.

42:57

>> Okay. So it sounds like you have you

42:59

started with the pipeline which I can

43:01

imagine first of all a lot of people on

43:04

this podcast talk about they start with

43:05

the pipeline. You have confidence it's

43:07

going to work. You control more of the

43:08

process. It's a little more

43:10

deterministic. I can imagine in this use

43:13

case though for the customer it feels

43:16

like they're on rails. It's not a like

43:20

open conversation where anything goes.

43:23

It's like the agent is guiding them

43:25

through their order taking. Whereas one

43:28

benefit I could see you getting from

43:30

shifting to a agent plus tools

43:31

architecture is the customer can drive

43:34

the conversation a little bit more. Is

43:36

that what you found?

43:38

>> Yes. So

43:41

we started using tools in general like

43:43

at the first moment. So we consider MCP

43:47

and workflows such as I don't know other

43:50

workflows with pipeline but tools was

43:53

the no-brainer for me and to get started

43:55

because we did the previous analysis and

43:57

then we saw that it was the fastest now

44:00

I think that's more what you're

44:02

mentioning is more of an emergence an

44:04

emergent property of the fact that we

44:08

decided tools this was already happening

44:10

right the agent has access to all these

44:13

tools at some point I did we did think

44:16

about using state just giving state to

44:19

the agent so that it knows okay now

44:21

right now it's just receiving the a the

44:24

or greeting the user now it's building

44:26

the order okay now the order is built so

44:28

the problem with that is that it would

44:30

happen what you just said they would you

44:31

would be having the customer on rails

44:34

and not giving them that freedom of

44:36

picking hey wait so actually there's

44:38

actually something a workflow that I can

44:40

mention that gives a good example so

44:43

what you can chat with the agent, start

44:45

building the order, you're done with the

44:47

order, and the agent sends you the

44:49

payment link. Before you even make that

44:52

payment, you can ask the agent to make a

44:54

modification to the order. So, they can

44:56

go back, update the order, and they will

44:58

send you a new payment link. You're not

45:00

on rails. You're free to to talk as you

45:02

would with a person on a call center,

45:05

for example, to take your order. So, I

45:07

think that's we were lucky enough to

45:09

make the right call at the beginning

45:12

>> early on

45:12

>> for tools. Yeah.

45:14

>> Okay. So, it sounds like so your agent

45:17

I'm imagining some of the tools that you

45:19

I'm not going to guess. Tell me some of

45:20

the tools that your agent has access to.

45:23

>> It's just built-in tools that we built.

45:25

So, basically the tools are for example

45:27

add products. You have a tool that is

45:30

basically create an order, add products

45:32

to the order. This tool with add

45:34

products to the order basically has all

45:36

the logic for adding modifiers,

45:38

comments, variations, etc.

45:41

You got tech product availability,

45:44

right? So you got search product, search

45:46

products. You got search knowledge base.

45:50

Of course, we have a knowledge base that

45:51

helps a lot about extra information. We

45:54

got geoloccation tools that helps hey

45:57

given this address figure out whether

45:59

you can use it. We got generate payment

46:01

link tool which basically takes care and

46:04

internally on all of these we have like

46:06

multiple providers. So depending on the

46:08

venue's configuration, if they have one

46:11

payment provider or the other, the agent

46:13

is agnostic to that. It just goes

46:15

through it and then the tool takes care

46:16

of that. That's the way Google thing

46:19

about tools, right? The same happens for

46:22

the geoloccation. If you most of the

46:23

times we use Google places API, but you

46:26

could use other one if you wanted. It's

46:28

kind of it we don't we try this is a

46:31

constant back and forth we have with

46:32

Santi most of the times. There's a lot

46:35

of things that go into the prompt, but

46:37

there's also a lot of things that should

46:38

be systematically happening, right? So

46:41

when the agent tries to take an action,

46:44

the tool should tell it whether that

46:46

action is successful or not and why so

46:49

that the agent can understand what's

46:50

going on. And this is like I was saying

46:52

the back and forth with Santi because

46:53

Santi takes a lot of time and working on

46:56

super amazing prompts that make the

46:59

agent talk like a real person. But then

47:02

we have the okay does the agent do this?

47:06

That's when the systematic

47:07

implementation has to come in. So that's

47:09

part of where we are mixing with

47:11

something where okay it's prompting but

47:14

the prompting should be related to what

47:16

the venues configuration is all about.

47:18

So we do we have a prompt composer

47:22

framework that we implemented which

47:24

inject fragments and depending on the

47:27

configuration it's just one fragment or

47:28

the other then so we have we do it

47:31

constant reviews about whether the

47:34

current prompt for how the agent should

47:36

reply con contains any logic that should

47:39

actually be implemented on the system

47:42

side or it's okay to implement it on the

47:44

more on the human side or how it should

47:47

reply. So yeah, a lot of pieces in

47:50

there.

47:50

>> Yeah. And many times we do MVPs for

47:54

different features through the prompt

47:56

and then we we build the features once

48:00

we realize that it's that it's helpful

48:02

and really needed by customers. We

48:04

assume the technical debt h to to test

48:07

if that's something that the customers

48:08

really need it.

48:10

>> Yeah.

48:10

>> Yeah. I love this. This is something

48:12

that the company every has written

48:13

about. Do you guys familiar with the

48:15

company every? No,

48:17

>> they're not. Oh,

48:18

>> okay. They are they're they build

48:21

themselves as the company of the future.

48:23

So, they have eight different products.

48:25

Each product's built by one person and

48:27

they're relying on AI. And they

48:29

introduce this idea of like agent first

48:32

development. And I don't mean coding

48:34

agents writing code first development,

48:36

but like you let the agent do the

48:40

feature. And then when you see if it's

48:42

working, you figure out, okay, what

48:45

parts of this should be deterministic?

48:46

How do we support this in code? How do

48:48

we optimize it?

48:49

>> But you're really relying on the agent

48:52

>> as your MVP to figure out does do

48:55

customers even want this? So, it sounds

48:56

like you guys stumbled on a very similar

48:58

idea, which is very cool.

49:00

>> Yeah, I have a concept that I've been

49:02

going through back and forth in my head.

49:04

I'm still trying to grasping into it,

49:06

but I'm starting to see ourselves as

49:10

building these products through how

49:12

there's business driven design or

49:14

business driven development, test-driven

49:15

design or test-driven development.

49:17

>> I think we're on the path of becoming a

49:21

company that that does conversational

49:24

driven design

49:25

>> because it's all about so this

49:27

conversation has to have this outcome

49:30

and it's not like the system or this

49:32

test. It's basically the conversation is

49:34

what triggers everything. So work on

49:36

that.

49:37

>> Yeah, I like that. I want to go back to

49:39

just your tool architecture and there's

49:42

something you said that I think is

49:44

really innovative that I want to dig

49:45

into which is your agent has tools and I

49:49

sounds like you're both at the prompt

49:51

level. You're injecting like I'm

49:53

assuming your agent needs to know for

49:55

this company this is what an order looks

49:57

like. These are the required attributes.

49:59

I can imagine that could happen at the

50:01

prompt. It could happen in the tool.

50:03

Just making sure that the agent has the

50:05

right stuff. I can see how this could be

50:08

a very, it's weird saying traditional

50:09

for a technology that's three years old,

50:11

but like standard agent turnbased.

50:15

You get a user message, you call a tool,

50:17

you do your thing, away you go. But

50:20

Juan, there's something you said that I

50:22

thought was really clever. You're

50:24

interjecting before the agent does a

50:26

tool call. You're looking at the user

50:27

message and trying to guess. Yes,

50:30

>> I want to give the agent more data so

50:34

that it doesn't have to do a tool call.

50:36

>> Say a little bit more about what how

50:38

you're anticipating there and what's

50:40

happening there.

50:42

>> So I want to give credit where credit is

50:45

used. So this was an idea that Santi

50:47

brought. So we just give it the message

50:49

of the user to the prompt and have it

50:52

understand that this that of course the

50:56

idea was amazing but it failed if we

50:58

just injected the prompt. So I had to

51:00

figure out a way to to do that and so

51:03

what we're doing is basically we have

51:06

smaller agents that are super fast

51:09

basically reading either just the

51:11

message of the person or like the last

51:14

conversations. And for example, we have

51:17

one super quick agent that what it does

51:19

is builds product queries for searching

51:23

products. So basically figures out

51:25

whether there's a product being asked on

51:27

the message and from that it builds okay

51:30

let's search for this product and so

51:32

while other stuff is happening it goes

51:34

and queries creates that query goes and

51:37

search and pulls that information.

51:40

Something similar happens with knowledge

51:42

base. So for example, instead of we're

51:44

to avoid having the agent call knowledge

51:47

base a lot, we grab the message of the

51:50

person and we just basically do a search

51:52

on the knowledge base just to see if a

51:53

one shot uh a fire just a single shot

51:57

gives us a result of knowledge base and

51:59

we also inject that. So we try yeah we

52:01

try to build that. Sometimes it's huge

52:03

and sometimes there's a lot of

52:05

information that is not relevant at all

52:07

which is the hard part because the agent

52:09

starts getting mixed up mixed up. But we

52:12

we are looking we still looking and we

52:15

have solved a lot of those but we're

52:16

still looking for so to improve that.

52:19

>> But that's how we're doing it even we

52:21

always on that session we inject the

52:24

state of the order for example. So it

52:26

knows a lot of things

52:28

without even needing to go through the

52:30

through tool calling to refresh its own

52:33

memory again and again. So like it's

52:35

like

52:35

>> we build temporary working memory on the

52:38

spot instead of having it processed. Of

52:41

course we have memory implementations

52:44

for for specific some the current

52:46

episode the whole session or even the

52:48

just the whole person

52:50

>> where we do know what the personnel

52:52

orders and kind of do suggestions or we

52:55

know who the person is. But in the end

52:57

yeah that's the strategy we use for now.

53:00

This is I think one of the most fun

53:02

parts of building AI products which is

53:04

just this like I to me it's bigger than

53:08

just context engineering but it's like

53:11

what goes in the prompt what data is

53:13

most relevant at this moment in time

53:15

even what tools should be available at

53:17

this moment in time and almost thinking

53:19

about your agent as a person and like

53:22

the way that we would train a person is

53:24

we don't tell them everything all at

53:26

once. We really look at based on the

53:28

conversation, what's the relevant

53:31

information for you right now? Like how

53:33

do we constrain your space so that

53:34

you're more likely to do a good job. I

53:37

could dig into this for the next hour,

53:38

which we probably don't have time for,

53:40

but it sounds like you do keep a profile

53:43

about your customer, so there's context

53:45

there. I imagine there's a lot of

53:47

context around here's what here's this

53:49

restaurant's menu. Here's all their

53:51

business rules about what goes with

53:53

what. on top of all your the goal is to

53:56

build an order and what does an order

53:58

look like for this company? Just tell me

54:01

a little bit about you must have had to

54:03

learn a lot just to make all those

54:06

pieces work together. What tell me a

54:08

little bit just orchestrating all of

54:09

that.

54:11

>> I think I have to tell you a little bit

54:12

about myself. So, as I mentioned, I've

54:15

been a developer since I was seven years

54:17

old. My degree is on music, not on

54:22

computer science. And that's because I

54:25

just learned to program and I learned

54:29

engineering just because I liked it and

54:32

I'm a self-taught engineer. So what that

54:37

gave me is not only I didn't only learn

54:40

how to build products or how to build

54:42

programs, how to correctly engineer or

54:44

architecture stuff, but also taught me

54:47

how to quick learn by just doing right.

54:51

And so in this scenario, I always been

54:54

like, let's do let's test if this works

54:57

and let's validate and then see or come

55:00

up with an idea just quickly see if

55:02

that's something that can happen. So for

55:04

example, this idea of having that system

55:06

prompt that immediate system prompt that

55:08

I was mentioning like I said we have two

55:09

system prompts. We have the main system

55:11

prompt which is what people mostly use

55:13

but then we have this smaller immediate

55:16

system prompt. And I came up with that

55:19

idea, but I wasn't really sure whether

55:21

that was supported by models. So I had

55:23

to go and search, hey,

55:25

>> would the model allow me to have two

55:28

system prompts like a main one at the

55:30

top and then one immediate injected and

55:33

then on the history of the chat that

55:35

system prompt wouldn't show up anymore.

55:37

It's just always appends itself on the

55:39

last message. Is that something that's

55:41

supported by model? So I had to go

55:42

through and learn about that. But it's

55:45

not that all the knowledge I had for how

55:47

models worked was already there because

55:51

I'm always like see okay this new thing

55:54

came out how do you use it? I already

55:56

knew it even before we can use it. One

55:59

quick example I want to give on that is

56:01

just for example I'm not sure if you

56:03

heard but just recently the second

56:05

version of Mercury came out by

56:07

Inception. Have you heard of Inception?

56:09

>> Yeah. this diffusion large language

56:11

model and I've been on that since their

56:14

first vision. I was like, man, this is

56:16

really something interesting. I think

56:19

this is really the path because it has

56:20

this uh this property of

56:24

being able to correct itself like

56:27

previous tokens can be corrected because

56:29

of just how it works instead of having

56:31

to you write a token and that's it. The

56:34

token is there. So learning all this

56:36

stuff gives you like the advantage of

56:38

already knowing what's possible and when

56:40

not. So when you have a problem you

56:43

already have the knowledge of the

56:45

different technologies on whether of

56:47

what like your tool set. is basically

56:49

the tools that you have at disposal.

56:51

Right?

56:52

>> What you're describing is selfishly why

56:55

I started this podcast.

56:57

>> Amazing.

56:58

>> Right. My thinking was if I collect all

57:00

these story I saw a need. I would see

57:02

lots of teams trying to learn this but

57:04

selfishly I was like if I interview a

57:07

bunch of teams about how they build AI

57:08

products by the time I'm building my AI

57:11

products I'll have heard a lots of ways

57:13

for how people have already solved the

57:15

same problems. That's a great way to do

57:17

it,

57:18

>> which is really fun.

57:20

>> That's great to see what you build.

57:22

>> Yeah. Yeah. I mean, yeah. I have several

57:24

AI products in the market now, which is

57:26

really fun. All around discovery

57:28

coaching.

57:30

>> Okay. Let's get into something you said

57:32

earlier, which is you told me about this

57:35

moment of doubt. This is really hard.

57:37

Can we do this? It wasn't good enough.

57:40

which really raises the question of how

57:43

are you evaluating if your agent is

57:46

working and I think with orders in

57:48

particular I could imagine some failure

57:51

modes that are really catastrophic. So

57:53

what are you doing to make sure this

57:55

works well?

57:56

>> We like to stick to one KPI which is how

58:00

many items did we identify correctly.

58:04

That is our most important metric to

58:08

really understand how well our agent is

58:10

working. Then it can say strange things

58:14

or it can use not the perfect

58:16

vocabulary. All of those things can be

58:19

improved. But the main thing main KPI,

58:22

how many of the items did it get

58:24

corrected?

58:25

That's it.

58:27

>> Yeah, I love the simplicity of that. And

58:29

I think from a customer standpoint,

58:30

that's probably what they primarily care

58:32

about. That's the only That's the only

58:34

thing they care about. Yeah.

58:35

>> Do you see like in your conversations,

58:38

do you get data to measure that? If you

58:41

mentioned at the end of a order, they

58:43

get a payment link, they pay for it, the

58:45

food gets delivered, they get a text

58:46

saying it's delivered. If something's

58:48

wrong, are they adding that to the

58:51

WhatsApp chat chat? Are they saying I

58:53

got the wrong thing?

58:55

>> I don't think if it ever happened,

58:58

maybe.

58:59

Yeah. I I think that the whole picture

59:01

here is that the agent can actually take

59:06

how do you say reclamosanti

59:08

>> claims

59:09

>> claims yeah so the agent can take claims

59:11

from the customer when there's a problem

59:14

and it even sends an email to the venue

59:17

so that they know about the problem if

59:20

there was a problem or how it was

59:21

delivered they even if this the user

59:23

sends a picture they can send it the

59:25

email will come with all the attachments

59:27

needed if there's an error in an order

59:29

it's more of a post post sale service or

59:33

customer service or that okay let's see

59:35

how we can solve this problem but to the

59:38

point that Santi was mentioning what we

59:41

tried to do is previously validate it

59:43

and Santi built an amazing tool and this

59:45

is amazing he just built this with

59:47

loable just a quick mention Santi is the

59:51

number one user of lovable in Argentina

59:54

in South America Latin America Santi

59:55

something like that

59:56

>> no in Argentina

59:57

>> okay within the 1% of users so They

59:59

reached out to him to say ask him but he

1:00:01

built this amazing tool which is mostly

1:00:03

front end of course there's a lot of

1:00:05

things happening but basically this tool

1:00:06

what it does it just uh he came up with

1:00:10

a way I mean something you want to talk

1:00:11

about it

1:00:13

>> while testing it we realized that

1:00:17

if we want to move fast you want to go

1:00:20

fast there are some things that can go

1:00:22

wrong of course especially with with AI

1:00:26

so we said okay there's actually two

1:00:28

things that we need to do because our

1:00:30

main goal again is to take the orders

1:00:33

correctly all of the items correctly. So

1:00:36

we can do that in two ways. The first

1:00:38

one is safe that works every time which

1:00:41

is human takeover. If we we do audits of

1:00:45

the live audits of the conversations. So

1:00:48

whenever we find things that are off, we

1:00:50

take over and we correct them by hand

1:00:53

and then we automate that. We fix that

1:00:55

problem and we automate. But before we

1:00:59

even get any of the agents into the

1:01:03

customer's hands, we do thousands of

1:01:06

tests. How do we do that? We actually

1:01:09

trained an agent that acts as a customer

1:01:13

to test the agent. We run literally

1:01:16

thousands of those during a night. uh

1:01:20

when they are done when one conversation

1:01:22

is done there's another agent that

1:01:24

analyzes the conversation and checks if

1:01:27

the order uh was correct if all of the

1:01:30

items were correct h and then after x

1:01:34

number of runs we have another agent

1:01:36

that analyzes wherever there was an

1:01:39

error right

1:01:41

and we start fixing that of course the

1:01:43

first time we did it we had a huge error

1:01:46

rate huge error rates but then we

1:01:48

started improving doing each of the

1:01:50

things that were happening. And now,

1:01:52

honestly, our production products might

1:01:54

make a few mistakes, but then we fix it

1:01:57

by hand if it ever happens. Once it

1:01:59

happened that we didn't catch the error,

1:02:02

uh, and we added an address that was not

1:02:06

the right one, one for delivery.

1:02:09

But since the customer gets a

1:02:11

confirmation ticket, he saw that the

1:02:13

address was not right and he called the

1:02:15

restaurant. So it actually was quite an

1:02:18

easy fix. Uh but there's a whole bunch

1:02:22

of agents testing the customer agents so

1:02:26

that the items are understood perfectly.

1:02:30

>> I also realized there like in my air

1:02:33

that I example that I gave I focused on

1:02:35

the wrong food getting delivered. That

1:02:38

was because of my bias with Door Dash.

1:02:40

feel like my experience with Door Dash

1:02:42

is I often get food from a totally

1:02:44

different restaurant that I ordered

1:02:45

from. It's just a weird experience.

1:02:47

>> Yeah.

1:02:48

>> Um, but I realized in your case,

1:02:50

>> you send them a payment link and I'm

1:02:52

assuming they get to see their order on

1:02:54

that page.

1:02:55

>> So, there's this like human in the loop

1:02:58

before the order is even finalized. Is

1:03:00

that true?

1:03:04

>> Yes. But that human is a customer and he

1:03:07

assumes that everything is okay. So they

1:03:09

are not testing it. They have

1:03:10

>> Okay. So you're not relying on that as a

1:03:12

feedback as a human in the loop.

1:03:13

>> No, no, no. The human in the loop is

1:03:15

someone at our team that jumps in case

1:03:17

there are any errors.

1:03:19

>> And how do you literally have somebody

1:03:22

monitoring all the conversations? Like

1:03:24

how are you detecting errors real time

1:03:27

>> for when we start with a new customer?

1:03:30

We try to audit them

1:03:33

>> all.

1:03:34

>> Okay. We have uh we have team members,

1:03:36

we have freelancers that helps us with

1:03:38

that. We do it ourselves as you might

1:03:40

imagine at 12 a.m. or 11:00 a.m. in the

1:03:44

night. We are reviewing conversations

1:03:46

many times.

1:03:48

But after finding out that was super

1:03:52

painful, we also have an agent that does

1:03:56

a revision automatically. Okay. and send

1:03:58

us an email, an alert email in case

1:04:01

there's any anything that needs to that

1:04:03

that that requires her attention.

1:04:05

>> So, this is part of onboarding a

1:04:08

customer,

1:04:09

you go through this testing phase to

1:04:12

make sure, correct,

1:04:13

>> the agent is interpreting the menu

1:04:14

correctly. It knows how to construct an

1:04:16

order. It's not something that you have

1:04:19

to do indefinitely. It's just part of

1:04:22

the like fine-tuning.

1:04:23

>> Just a few weeks,

1:04:25

>> correct? Just a few weeks of on boarding

1:04:27

and then it just starts on its own.

1:04:29

Yeah.

1:04:30

>> Yeah. It used to be three months and as

1:04:32

we improve this the time for boarding is

1:04:35

reducing which is basically our biggest

1:04:37

challenge and what we're working on is

1:04:38

like improving the on boarding time to

1:04:40

make it super super fast. Like we think

1:04:43

we've already solved the messaging part.

1:04:45

We I mean even today when we are

1:04:48

reviewing calls messages with Santi just

1:04:51

so you know the idea it's like every

1:04:54

noon or evening where like lunch or

1:04:58

dinner time Santi and I are sitting on

1:05:00

our computers just watching these

1:05:02

conversations going in orders being

1:05:03

placed and just seeing them go and it's

1:05:06

like we message each other going did you

1:05:09

just see what it how it solved the

1:05:11

problem is we are even are amazed about

1:05:13

how it's working on so Yeah,

1:05:15

>> it's impressive. So, we think that part

1:05:18

is solved. And the part that we're

1:05:20

trying to solve right now is basically

1:05:21

decreasing the onboarding times.

1:05:24

>> Times when the type of restaurant is

1:05:26

new, you might have a longer on boarding

1:05:28

because of all of the different

1:05:29

products, but for example, you can get

1:05:32

us any pizzeria and we will get it at

1:05:34

pizza stores and we will get it set up

1:05:36

pretty quickly because we know how that

1:05:39

business works. That's it's cool to see

1:05:41

like how you can build iterative domain

1:05:44

knowledge and start to reduce that

1:05:46

onboarding time for different types of

1:05:48

businesses. This is great. It's really

1:05:51

clear you're passionate about your

1:05:52

problem space and that you've really dug

1:05:54

in and that you have an equal passion

1:05:56

for the technology which is fun to see

1:05:58

too. Is there anything you wish I had

1:06:00

asked you that I didn't?

1:06:09

I think your questions were

1:06:11

amazing. We got to talk about a lot of

1:06:14

stuff.

1:06:16

>> All right, then let me ask you one last

1:06:18

question before we wrap up. What's next?

1:06:20

What's the big challenge that you're

1:06:23

tackling next?

1:06:25

>> Yeah, it's a great question and honestly

1:06:28

super proud of our product and we've

1:06:31

seen it working in lots of venues. So

1:06:35

now our goal is to scale it. We want to

1:06:38

be we want restaurants all over the

1:06:41

world to be able to use this tool. We

1:06:43

are focusing on Argentina, Mexico, USA,

1:06:46

and Spain. But we really believe that

1:06:49

this we're already seeing the numbers

1:06:52

and it's impressive how much it helps

1:06:55

businesses that that can use these type

1:06:58

of tools. So next step, let's get let's

1:07:01

put let's get it out there. Amazing.

1:07:04

Hopefully this episode will help get it

1:07:06

out there. Santi and Juan, it's been

1:07:09

super delightful to hear about your

1:07:11

story and to learn about your product. I

1:07:13

look forward to when I get to actually

1:07:15

order a meal through WhatsApp. So, I

1:07:18

will keep an eye out for it here in the

1:07:20

US.

1:07:21

>> You will. Thanks so much for having us.

1:07:24

Group super good questions.

1:07:25

>> Super super nice. If you enjoyed this

1:07:28

conversation, please subscribe in your

1:07:30

favorite podcast app and give us a

1:07:32

rating as it helps others find the show.

1:07:34

Thanks. I appreciate it.

More transcripts

Explore other videos transcribed with YouTLDR.

Get the TLDR of any YouTube video

Transcribe, summarize, and repurpose videos in 125+ languages — free, no signup required.

Try YouTLDR Free