Latest / Agentic Conversations (formally mlops.community) / Building MCP Before MCP Existed: Inside Despegar's Sofia Agent

Transcript

Building MCP Before MCP Existed: Inside Despegar's Sofia Agent · Agentic Conversations (formally mlops.community)

• 0:00
  All right, we've got a special episode where I talk with
• 0:02
  Nicholas and AIPM from the company Despegar in Latin
• 0:08
  America, all about how they are building their concierge agent
• 0:13
  for travel called Sofia. We get into many different
• 0:18
  things, even how they built what is a competitor to MCP before
• 0:25
  MCP was even a standard and about some of the key challenges
• 0:30
  they faced while putting this AI agent, Sophia into production.
• 0:36
  Let's listen to it now. In our multi agent, we have one
• 0:40
  application that's named Chappie and it's like the brain of
• 0:43
  Sophia. Each agent, it's like kind of a,
• 0:46
  we call it a category or a different flow, for example,
• 0:49
  flights, hotels, activities, cars, etcetera.
• 1:01
  It's not like we are like 20-30 people developing Sofia, but
• 1:06
  anyone in the company can create a flow obviously with our OK,
• 1:12
  let's say, but with our supervision, but anyone in the
• 1:15
  company can create their own flow and work on it.
• 1:19
  And because of the know how, like we have some know how, for
• 1:23
  example, how on how to sell hotels.
• 1:26
  But the the best team to do that is the hotels team so they can
• 1:32
  take their flow and make the changes they think the they
• 1:36
  think they that has to be made. Someone who wants to build a
• 1:41
  flow is expected to. To be technical, yeah.
• 1:44
  OK. Like in, in, in this regard, we
• 1:47
  are organized as a squads and each of each squad, let's say
• 1:51
  the hotel to go back to that example, has a product member
• 1:55
  and IT team or developer team and a user experience team as
• 2:00
  well. So they took their flow and make
• 2:05
  the changes and improvements that they think are the best.
• 2:08
  Yeah. And it makes sense that you
• 2:09
  would want to spread out that work as opposed to 1 centralized
• 2:14
  team that's trying to develop everything versus just give the
• 2:20
  foundational building blocks that folks can then customize on
• 2:24
  top of the. The most of the categories are
• 2:26
  the the flows agents that we build.
• 2:29
  We'll start building it and then we pass that to the
• 2:33
  corresponding team. So you, you think maybe we
• 2:37
  should do something along these lines and then you build like
• 2:40
  the basics and then they can customize?
• 2:43
  Yes. And the whole idea there is it's
• 2:48
  all based around chat. It's all based around chat.
• 2:51
  Also boys like you can call in our after sales team.
• 2:55
  If you call our the speaker number, Sophia will will will
• 3:01
  contact you and then you can you can be routed to an agent, but
• 3:05
  Sophia will be the one that that first call first talk to you.
• 3:09
  Yeah, because I remember seeing that demo where it was like,
• 3:12
  hey, what's my flight number? What's my when do I check in
• 3:16
  simple questions that somebody maybe is is calling because they
• 3:20
  don't really know and it takes a lot of work.
• 3:22
  It would surprise you like the simple of pay or basic questions
• 3:26
  that we have in after sales like which is my checking number,
• 3:30
  where's my voucher, where's my invoice?
• 3:34
  And it was like that simple questions are perfect for
• 3:39
  Sophia. And if the person has, I don't
• 3:41
  know like a reschedule of the flight, probably they will, we
• 3:45
  will route you to an agent. But the workload of the agents
• 3:50
  were like they are. They are more optimized to do
• 3:54
  what? It's a complex question.
• 3:56
  Yeah, I remember my buddy Elliott, who works at Boca
• 4:00
  Health, and he does stuff for doctors and clinicians, and it's
• 4:06
  more or less the same thing, but all around voice.
• 4:09
  And so people will call a doctor's office and say things
• 4:13
  like, how long was I supposed to fast before my surgery tomorrow
• 4:17
  or whatever it was? And he said that like 80% of the
• 4:21
  questions were super simple, just basic things that you
• 4:26
  probably got on some form or some e-mail.
• 4:29
  Yeah, in a document, but you don't want to look for it, which
• 4:32
  I find hilarious. Like if you just Googled it you
• 4:34
  could probably find it, but it takes a little.
• 4:37
  Bit that your doctor? Says that, yeah.
• 4:40
  And it probably takes a little bit more effort to go and click
• 4:45
  through the UI and figure out. And then I do think, and you've
• 4:50
  probably thought about this a lot as the product manager,
• 4:53
  you're leaving a little bit of effort or money on the table.
• 4:58
  Because in your use case yesterday where someone called
• 5:03
  and they said, what's my check in or when can I check in,
• 5:05
  right. Yeah, I want to be able to say,
• 5:07
  OK, go check me in. Yeah, like in in the after sales
• 5:12
  part, what we are building right now, it's like going through all
• 5:17
  the process. For example, if you have a a
• 5:20
  change in your flight like we are trying to Sofia be able to
• 5:26
  do the task for you, we are not there yet.
• 5:29
  We are starting with that. So we complete like we have two
• 5:33
  like 2 frameworks or two visions, one to, to grow
• 5:38
  horizontally and one to grow like deeper in, in each flow.
• 5:43
  So we have cover horizontally every like almost everything and
• 5:47
  we are now building like for each one of the topic deeper
• 5:52
  solutions. So you don't have to contact or
• 5:55
  you don't have to do it yourself any of the OR any of that task.
• 5:59
  Yeah, that's very special. It got me thinking as I was
• 6:02
  reflecting on that use case and wanting to be able to just say,
• 6:06
  OK, now go check me in. Then I was like, wait a minute,
• 6:09
  why do we even have to check into planes anymore?
• 6:12
  I don't know. I took.
• 6:13
  That's a good question. I was like what the fuck is this
• 6:16
  about we're getting psyopped 100% right now.
• 6:19
  Like what is the check in even for?
• 6:21
  I could see what it was for back in the day.
• 6:23
  Yet but now like it doesn't make any sense.
• 6:26
  Right. Am I?
• 6:27
  Am I crazy? Here.
• 6:28
  No, I'm with you. I'm with you on that.
• 6:30
  I imagine there's a lot of difficulty though with the
• 6:36
  identity or being able to vouch for like Sophia checking in.
• 6:40
  For me, the thing that is difficult is knowing that I'm
• 6:45
  actually telling Sophia. And you are you and the
• 6:49
  documents and etcetera. But then again, like, why would
• 6:51
  I tell you to check in if I if it wasn't me?
• 6:54
  Like, it's just, it's a ridiculous thing the whole check
• 6:56
  in. Yeah, I go back to that.
• 6:58
  I gave a few talks about this experience when I was traveling
• 7:02
  to Portugal from Germany and my flight was 3 hours delayed.
• 7:08
  And in Europe if it's three hours delayed you are able to
• 7:11
  get money back. And I was going back and forth
• 7:13
  with my LLM of choice right and saying like hey my flight's
• 7:17
  delayed this much, here's the flight number, blah blah.
• 7:19
  Do I am I entitled to any money? And I go through that whole
• 7:23
  thing and it's like, yeah, you can get 250 bucks per passenger
• 7:27
  and I was traveling with my wife and two daughters and I'm like,
• 7:30
  shit, I am rich. This flight like the.
• 7:33
  Getting rich, yeah. Just paid for my vacation and
• 7:37
  what do you think I tried to do that had me cut the
• 7:41
  conversation? What was it that made me realize
• 7:43
  like, ah shit, I'm not actually going to get the money?
• 7:47
  I don't know. I tried to say OK, go and make
• 7:51
  the claim for me. OK.
• 7:53
  And it. Didn't it was like I'm an AI
• 7:55
  assistant, I can't do that. I'm like, well, you're not
• 7:57
  assistant much, are you? You're basically just giving
• 8:00
  some random information and. I'm giving giving you tasks to
• 8:04
  do yourself. You're my imaginary friend who
• 8:08
  now is increasing my workload. Yeah, yeah.
• 8:12
  And so that is a perfect use case.
• 8:14
  Like I would love it if I could just get file claims through a
• 8:18
  chat. Bot through a chat bot.
• 8:19
  But not, and this is where I like in my talk, I was saying
• 8:23
  there is this really weird paradigm that's happening
• 8:26
  because all of there's a lot of Google services or if you Google
• 8:32
  file my claim for me, you can find pages and pages of websites
• 8:37
  that will file your claim. But it's not inside of the
• 8:41
  native chat bot that I want to have.
• 8:43
  Maybe they have their own chat bot, but it's not.
• 8:46
  Like the one that you use. Yeah.
• 8:48
  And it's not like I can just say, oh, hey, Gemini or Chad GBT
• 8:51
  or Claude, like please use this for me.
• 8:54
  Yeah, use this tool and file it for me.
• 8:57
  And so that was where me as a customer, I had so much
• 9:01
  friction. And if I was already on Sofia,
• 9:04
  I'm already interacting with my all of my travel itinerary
• 9:08
  through you in despair guard. It would make so much sense.
• 9:12
  That you have all covered there. Yeah, just file that claim.
• 9:15
  They did this. And if it was easy, I would
• 9:19
  probably be making a lot more money because there's so many
• 9:22
  times that my flight's been late and I'm just like, Nah, like I
• 9:25
  don't care that much. The airplane services do
• 9:30
  anything they can to like get out of filing it or get out of
• 9:34
  like paying it. And so, you know, there's going
• 9:37
  to be a back and forth and you know it's going to take up a lot
• 9:40
  of time when you do it. But that seems like a perfect
• 9:42
  use case for. An agent for any for Sophia for
• 9:45
  an AI use yes. Yeah, so I'm giving you all
• 9:48
  kinds of ideas. Here what I'm going to take it.
• 9:51
  I'm going to take the ideas. What are you actually working
• 9:54
  on? Because I'm just giving you
• 9:55
  these utopian scenarios. Like we have like one main focus
• 9:59
  in Sophia right now. It's like make our users more
• 10:05
  awareness of Sophia. Like we have like 2K PIS that we
• 10:10
  follow a lot. The first one is how many of our
• 10:14
  clients interact with Sophia. And of course then we double
• 10:18
  click on how many times or how useful Sophia was.
• 10:22
  But we have 3030% of our clients interacting with Sophia in any
• 10:27
  part of the journey. And we also like follow a lot
• 10:33
  the KPI of Sofia's gross booking.
• 10:35
  That's like says like saying that Sofia sales like in how
• 10:39
  many sales Sofia was a a very important part of the customer
• 10:45
  journey. How are you?
• 10:47
  Presenting Sofia, Yeah. Exactly.
• 10:49
  We have like we have two channels in in the sales part.
• 10:53
  The first one is or yeah, the first one is WhatsApp.
• 10:56
  Like if you talk for to WhatsApp to the Espicard through through
• 11:00
  WhatsApp, Sophia will text you. Then it can route you to an
• 11:04
  agent. But but Sophia can offers you
• 11:07
  flights, hotels, itineraries, activities, any product that you
• 11:13
  think and also can help you with destinations with what to do in
• 11:19
  each in each in each part of your travel.
• 11:24
  The the value proposition of Sofia is that Sofia can offer
• 11:27
  you like real time offers, like it's not a cachetta that we
• 11:33
  found some we found in our documents or in our history.
• 11:36
  But Sofia make the actual search for you and give you the detail
• 11:40
  of the product that you click and you buy at the moment.
• 11:43
  So I think that that's one of the of our main value
• 11:48
  propositions. So you can through WhatsApp you
• 11:51
  can explore all the, let's say the shopping flow of flights,
• 11:56
  hotels, etcetera. So that's one of our channels
• 12:00
  and the other channel is in our app or in our website where
• 12:03
  Sofia has different placements around around the the platform.
• 12:09
  But since July, August, we are like trying to present Sofia to
• 12:16
  more and more customers like like we want to our customers to
• 12:21
  be more aware of Sofia in the sales part.
• 12:24
  And two main or one of the things that we do.
• 12:27
  It was in the search box where you complete your destiny, your
• 12:30
  your dates, your distribution, like how many are people are
• 12:34
  traveling. We like add a switch or a toggle
• 12:38
  down that down there and it says like you say hi with my search
• 12:43
  and when you click it, you go directly to Sofia and you go
• 12:47
  through all the shopping flow in Sofia and you only have to pay
• 12:50
  in our, let's say traditional flow.
• 12:53
  So that's one of the things that we are doing.
• 12:55
  And we are also adding, I don't know if anyone has seen the
• 12:58
  Spegar, but we have like in the, in the top of our platform, like
• 13:03
  the buttons of each products that we have.
• 13:06
  We are like adding now a new button.
• 13:09
  That's it's a model AI like AI mode where you have an open
• 13:13
  search box and you can search wherever you want.
• 13:15
  And Sofia will answer. So you don't have to be like a
• 13:19
  very study way of searching. Like it's the way that you used
• 13:23
  to search in travel like Buenos Aires to Madrid from this day to
• 13:28
  that day for two people. But you can say, hey, Sophia,
• 13:30
  give me the cheapest flight between these days and Sophia
• 13:35
  will help you with that request or Sophia give me the offers or
• 13:38
  how are the prices now to go to Madrid.
• 13:41
  So we are like enabling a new way so to the users to search
• 13:48
  for our products and for for their trips.
• 13:52
  Oh man, it is so cool to think about the idea of just like I
• 13:58
  have this block of time free, where should I go?
• 14:02
  Yeah, that's what Sophia is going to help you with.
• 14:04
  Let's say like. And I know Airbnb did that and I
• 14:08
  wonder how successful it was because I haven't heard anything
• 14:11
  since, but it was it was kind of like show me just unique places
• 14:17
  to go to and isn't. Before AI, it wasn't easy for us
• 14:22
  like an OTA to be part of the dreaming part, the dreaming
• 14:28
  section of your trip. Like where do I go?
• 14:30
  I have this time free. How do I plan this?
• 14:33
  Because we were very transactional.
• 14:35
  So Sophia, it's also covering that that, that we didn't in in
• 14:41
  the desperate platform, in most of the OTAs we didn't have
• 14:44
  cover. So it's like a very new use case
• 14:47
  for our platform and we are improving it like every day.
• 14:52
  Let's say like how we recommend you, yeah, the destinations for
• 14:56
  you, how many batches do you have?
• 14:59
  What are your preferences, what to do in each in each
• 15:03
  destination? So that's like a new part for
• 15:06
  our platform. I'm Sophia.
• 15:07
  It's it's the best place to do the best place to do it.
• 15:11
  See the different user segments? Because I imagine a subset of
• 15:14
  your users only uses it for flights, and then a subset only
• 15:18
  use it for hotels, and then there's some people that use it
• 15:21
  for the full end to end. Is that one of the KPIs you're
• 15:24
  looking at too? Is like, how many more things
• 15:26
  can we get them to engage with? Yeah, we'll, we'll look at that
• 15:30
  like we used to have one KPI that was like in a scope
• 15:34
  conversation like hey, Sophia can't answer the the the the
• 15:39
  user question or the user prompt.
• 15:42
  And that KPI was like 9596% like very asymptotic.
• 15:47
  So we are not tracking it anymore and we are now like for
• 15:52
  now we think we have covered most of the use cases that the
• 15:56
  customer request us. We have like two ways to keep
• 16:00
  growing Sofia. One, it's adding use cases that
• 16:04
  the customer didn't know they want, but they will use and we
• 16:08
  have some of them. And also improving the ones that
• 16:16
  we have today, like we can be better recommending your
• 16:19
  destinations, we can be better making you better trip plans,
• 16:25
  improving our flight. Like for every product.
• 16:31
  Our mantra is, for example, for the flight, for the flight
• 16:34
  shopping flow, we want Sophia to be able to perform things that
• 16:39
  you can do on a traditional page.
• 16:41
  Like if you go for the for the traditional platform and search
• 16:45
  for a flight, it's not easy to compare dates like one day
• 16:49
  prior, one day later. Or if it's Madrid versus
• 16:53
  Amsterdam. Exactly.
• 16:54
  So we are trying to focus Sophia on performing that things that
• 17:00
  it will be very difficult for a normal user to do in the
• 17:03
  traditional platform. So we add value.
• 17:06
  Various browser windows open to. Compare.
• 17:08
  That's what all we all do. We want Sophia like to be able
• 17:13
  to solve all of this in the chat.
• 17:16
  You know what, I I'll give you some more things that I want as
• 17:20
  a user. This is only because we had a
• 17:23
  hackathon in the summer and the winners of the hackathon came on
• 17:28
  the podcast and they talked to us about what they built and it
• 17:31
  was just the little chat bot that you would drop into your
• 17:34
  WhatsApp groups and then it would help you organize with
• 17:37
  friends. Yes, we are looking into that
• 17:41
  because of as we have Sophia in WhatsApp, it's the perfect use
• 17:44
  case like add Sophia to your group friends chat, chat and
• 17:49
  call Sophia wherever you want to check something like, hey, how's
• 17:53
  the OR how many do I have to pay to get a flight to Rio tomorrow?
• 17:58
  I don't know or on February. So it's one of the use cases
• 18:02
  that we are very. Yeah, I'm talking more, and you
• 18:05
  probably are too, in the dreaming phase.
• 18:08
  I've got a group of my friends that I used to go skiing with
• 18:12
  and we want to organize a ski trip, but we don't know when and
• 18:16
  we don't know where and we don't know where we're going to stay.
• 18:19
  And so Sophia, the way that I think these folks had done it is
• 18:24
  that you call the chat bot or the chat bot interviews you and
• 18:29
  then it takes down the dates, it takes down your price
• 18:31
  sensitivity, it takes down, you know, where, what kind of trip
• 18:35
  you want to go on because it's not always so clear like, hey,
• 18:38
  we want to go skiing. It's like, what are your top
• 18:41
  three things to do on holiday? And then it says, hey, I've got
• 18:45
  these different options for holiday.
• 18:47
  Whoever's in, give a thumbs up emoji and I'll buy your flight.
• 18:51
  OK. And that's the task for you as
• 18:53
  well. I think that one of the things
• 18:56
  that you mentioned and we have, we are always looking at this in
• 19:01
  the traveling industry like the customer doesn't care if you
• 19:05
  took three more seconds or if they ask you questions because
• 19:09
  it's going to be an experience that he's like waiting for.
• 19:13
  It's not like I'm asking any other LLM like, hey, what's my
• 19:18
  routine for today and it takes forever or ask me questions.
• 19:22
  But in the, in the travel industry, users are willing to
• 19:26
  give us that time to get an, a better answer, a better
• 19:29
  recommendation, the perfect products for for their trip.
• 19:33
  So we are, we have to be careful just not to ask too much
• 19:38
  question, but to explore that that that idea of asking user
• 19:44
  informations because they are willing to to give us that.
• 19:47
  Yeah, and if you make it frictionless or?
• 19:51
  With with any gamification, let's say like.
• 19:54
  Yeah, we're just OK, it's an emoji and then boom, it knows.
• 19:57
  That's that's my confirmation. It's not like I have to type
• 20:01
  anything or do anything in that regard.
• 20:03
  And the other thing that I was thinking is how you are able to
• 20:08
  do so much with that Direct Line of communication on WhatsApp
• 20:12
  because. Me, I'm a bit of a coffee snob
• 20:15
  and when I go to places I like to know where all the best cafes
• 20:18
  are around where I'm staying. And if you know that then you
• 20:22
  can send me a nice little report of hey, here's all the ones.
• 20:26
  But then even more than that you can say I created this shared
• 20:31
  Google Maps yes, and you just have to join my Google Maps and
• 20:35
  it has everywhere already marked.
• 20:38
  Bar mark with bins like where you have to go.
• 20:40
  Where you go, we are. Starting to explore like in trip
• 20:45
  use cases like we, we now have like Sofia outbound messages to
• 20:51
  our customers mostly in the pot sale journey of the of the of
• 20:58
  the of the user. But we are starting to explore
• 21:02
  what's or how can we deliver value to the customer in the
• 21:07
  entry part of of their journey. So yes, one that we were
• 21:13
  thinking was like restaurants recommendations on hey, I know
• 21:17
  that, I know the hotel that you are saying because you buy it
• 21:20
  with the cigar. So hey, this is this is the
• 21:22
  nicest place near you, this is the nicest restaurant etcetera.
• 21:27
  So we are starting to explore that new face that.
• 21:32
  Yeah, I remember my family came and visited me when one of the
• 21:36
  first years that I was living abroad and they expected me to
• 21:39
  be the tour guide and. Of course I would expect the
• 21:42
  same. Which I, we kind of made a joke
• 21:46
  of it because in English, you know, you call when you can't go
• 21:49
  somewhere and you have to go around.
• 21:52
  It's called the detour. OK.
• 21:54
  And I'm Demetrius. And so I was given a detour.
• 21:57
  It was basically everywhere we were going was a bit of a
• 22:00
  roundabout way to get there. And it was, it was a complete
• 22:03
  detour. But at the end of the day, what
• 22:06
  I put together was like a pretty legit guide on the place we were
• 22:12
  going, why we're going there, why I thought it was cool, the
• 22:15
  different restaurants that we had to choose from, that I think
• 22:20
  everybody would enjoy and that you can just create now.
• 22:25
  Yeah, I don't need to spend any time and each person can get
• 22:28
  their own. Personalized recommendation for
• 22:32
  their own preferences. Comes back to this idea that I
• 22:35
  was talking to Tiago about personalization at scale is so
• 22:40
  cool and the new ways of doing it.
• 22:44
  What I've been realizing talking to everybody here is it's almost
• 22:49
  like WhatsApp for the consumer is becoming the Control Center,
• 22:53
  just like Slack for the. Enterprise.
• 22:57
  Is becoming the Control Center right?
• 22:58
  Because WhatsApp, it's what it is.
• 23:00
  That's why we said, we said like I think it was an year ago,
• 23:03
  like, hey, we have to have Sofia in WhatsApp.
• 23:05
  We can't miss this opportunity. We have to work on this and we
• 23:10
  took the project and we developed Sofia in our existing
• 23:14
  lines in WhatsApp, so. Yeah, and for the Americans, I
• 23:17
  guess they're not seeing it as much, 'cause whenever I go to
• 23:20
  the US, it's always I. Message of.
• 23:22
  Course, and anywhere outside of the US.
• 23:26
  Yes, in America, what's up is very big.
• 23:29
  In, in Europe and India, all of that, it's so popular.
• 23:33
  And so having that as your Control Center where potentially
• 23:37
  I'm interacting with various agents.
• 23:38
  I've got my Despergar agent, I've got my I food agent in.
• 23:44
  If I'm in Brazil, I know I'm talking to all these.
• 23:47
  Agents and. Yeah, agents to get my stuff
• 23:49
  done. It will make your life easier.
• 23:51
  Yeah, yeah, yeah. Especially if you're able to get
• 23:57
  it to do more stuff. But then what I am constantly
• 24:01
  thinking about is I don't want to have to like switch between
• 24:04
  agents. Yeah, I just want to talk to
• 24:07
  Sophia and then Sophia can talk to the I Food agent.
• 24:11
  And so, yeah, it's like having an ecosystem of our agents all
• 24:15
  in one place so you don't have to take care of which one should
• 24:18
  I text, but we can solve it that for you, yeah.
• 24:22
  Yeah. You know, the more things you
• 24:24
  throw at it, it's going to be able to kind of be able to do
• 24:27
  it. And that I think is the current
• 24:31
  browser wars that are happening. You know how like in 2014 or
• 24:36
  2010, I can't remember exactly when, when Chrome was coming
• 24:40
  out, you had the new browsers. And that was like the browser
• 24:45
  wars because it was the front page of the Internet.
• 24:47
  It was how it was our gateway to.
• 24:49
  How the the platforms were presented to the customer?
• 24:52
  Yeah. And now it feels like that's the
• 24:54
  new gateway to the Internet. Yeah, Is.
• 24:59
  Yeah. And what chat interface we use?
• 25:03
  So and who's chat? Maybe it's not the interface as
• 25:08
  much as it is what agent we choose to use.
• 25:12
  And each agent at the end of the day is probably going to have
• 25:15
  their own chat I guess. Which like, it doesn't worry you
• 25:20
  that Meta might just say, oh actually everything like that we
• 25:25
  can do now. For us, it's very valuable to
• 25:28
  have our customer in our application.
• 25:31
  So we always try to to have that and to make them use our app.
• 25:39
  So Sofia, it's let's say more optimized in our app that in
• 25:43
  WhatsApp because of Whatsapp's the limitations of templates,
• 25:49
  flows, etcetera. So in our app, you have a richer
• 25:55
  UI Sofia than you have in WhatsApp because of the
• 25:59
  capabilities that that we can build.
• 26:02
  So I think that for now there's value on having using Sofia in
• 26:10
  the speaker website or the speaker app.
• 26:12
  And I like how you broke down yesterday, the different phases
• 26:16
  you have what you have the dreaming phase, which is when
• 26:18
  you're planning the trip and you're starting to think, where
• 26:20
  could I go? Yeah, we have the the dreaming
• 26:22
  phase. That is the first one that as
• 26:24
  you just said it, it's like where can I go?
• 26:26
  What do you recommend me? That's this is my preference is
• 26:28
  I have this budget. Then you enter when you decide
• 26:32
  your where do you want to go, you enter to the planning phase
• 26:35
  like when do I, when do I want to go?
• 26:39
  What can I do there? What are the main attractions?
• 26:44
  Where am I going to stay? Where you're going to stay and
• 26:47
  then you enter to like what we call in the booking part or the
• 26:51
  the purchase part when you select the right product for
• 26:56
  you. So we have those three important
• 27:00
  after purchase, before purchase parts and as I mentioned, we try
• 27:08
  to cover it all with Sophia that it's a a new use case for our
• 27:14
  platform. So we are giving users more
• 27:18
  value than we used to as test the guard.
• 27:22
  And then there's other phases of the trip.
• 27:24
  You have the actual trip. Itself before the actual trip,
• 27:29
  you have the anticipating parts like you buy a flight from us.
• 27:36
  But if I know, I don't know if I know that you are going to go to
• 27:39
  ski like I know that you have to that I can recommend you or I
• 27:44
  can help you with ski passes, equipment, etcetera.
• 27:48
  So that's all the anticipating part, like helping the custom,
• 27:52
  the the user to have all that he needs when he arrived at the
• 27:57
  destiny. Then we have the experiencing
• 27:59
  part that it's one that like in tribe, like we were talking
• 28:04
  before, like how we can deliver value when you are in your
• 28:07
  destiny. Let's say like, hey, this is a
• 28:10
  very great place to take coffee. And then we have the sharing
• 28:13
  part like we are, we are also starting to work on, on this,
• 28:18
  but on travel, it's very like useful to see what other
• 28:23
  customers or with other with other people than on that place
• 28:28
  or in that on their trip. So we think that it's going to
• 28:33
  be important for us to have like AI don't think like maybe like a
• 28:38
  community. But if you can't, if you're in
• 28:41
  your dream, let's say you want to go to ski.
• 28:43
  I have went to ski like 2 months ago.
• 28:46
  I have a upload to Sofia, my review, my, my recommendations,
• 28:51
  etcetera. And when you enter Sofia and you
• 28:54
  say, hey, I want to go ski, Sofia tells you, Hey, I have
• 28:58
  this 1000 reviews and for skiing users are recommending this.
• 29:04
  It will be very more powerful than and me telling you this
• 29:09
  like it's going to be more. Social Proof.
• 29:11
  Yeah, it's not social proof that we are going to to trying to
• 29:14
  have. Yeah, I was thinking about if
• 29:16
  you needed to actively reach out to customers to get those
• 29:21
  reviews or if they were just willingly putting them and
• 29:24
  saying, hey, I went here, it was great, I did this.
• 29:28
  I think that it's like a combination of those two things
• 29:31
  like we have to to go to get our customers to give us the review,
• 29:36
  but some of them also are doing it now without any.
• 29:41
  Yeah, which also feels like a place for Sophia if you want to
• 29:46
  get reviews like Sophia could call people, say, hey, you went
• 29:49
  on this trip. How did it go?
• 29:51
  Yeah, Tell me about it. Yeah.
• 29:53
  What did you like? What did you not like?
• 29:54
  I'm getting a lot of ideas here so.
• 29:57
  Dude, you might as well hire me at this point.
• 29:59
  Shit. You're making my job easier.
• 30:02
  I know what actually do you do because I'm doing your job.
• 30:05
  Right now I'm talking to you. That's my job now.
• 30:07
  You can just, I could be in a consultant if you want, I'll
• 30:10
  work part time. But the, the thing that, that I
• 30:14
  was also wondering about is when you're going on the trip, you
• 30:19
  have that anticipation phase. And there's certain times that
• 30:23
  me as a traveler, I will go some places, but I don't want to
• 30:30
  bring something or I forget something and then I just order
• 30:35
  it to my hotel, right? Sometimes it's, I order it in
• 30:38
  advance and this is when. So I'll give you a very concrete
• 30:44
  use case. OK, I'm going to Greece.
• 30:46
  I'm only taking a carry on, I can't take liquids and I'm going
• 30:53
  to sound so cheap right now but fucking sunscreen at any little
• 30:57
  kiosk in Greece is like €10 and if you order it.
• 31:01
  It's like 4. Yeah, no, it's like 2.
• 31:04
  So there's times where I've asked my godmother who lives in
• 31:11
  Athens to send to my hotel. She buys like a 2 year old thing
• 31:16
  of sunscreen, sends it to my hotel in whatever island that
• 31:20
  I'm going to so that I can have sunscreen, right?
• 31:24
  I'm thinking about that like whole customer journey.
• 31:26
  There's probably a million other things that people want when
• 31:28
  they get to the hotel that they want waiting for them.
• 31:31
  And if you said, hey, you're going to this ski trip, by the
• 31:35
  way, do you need a new ski helmet or?
• 31:38
  You have glasses. Yeah, when was the last time you
• 31:40
  got goggles? We got these new ones on offer
• 31:42
  in case you want to like look cool for your friends because
• 31:45
  everyone wants that external validation.
• 31:48
  Yeah, it's like like very near to your trip, like sending you,
• 31:53
  hey, this is the list as the list of things that you are
• 31:56
  going to need. Do you want like?
• 31:58
  Any of them. Any of them do you want any, any
• 32:01
  advice on any on any of them? So yeah, yeah, that's like very,
• 32:06
  it's like very near your trip, but it's very useful use case.
• 32:11
  Yeah. Then you just use OLX for you
• 32:16
  guys. The architecture that you
• 32:18
  decided on was Sophia's going to be the gateway and then you're
• 32:22
  going to, yeah, you can call tools or different agents that.
• 32:27
  The houses and their tools. Yeah.
• 32:29
  And they're very specific to those parts of the business.
• 32:33
  And so if it's a complaint, I can complain directly to
• 32:36
  Sophia's and then it goes to the customer service agent as
• 32:39
  opposed to a very verticalized agent that is only made for one
• 32:45
  thing. And if now I complain to that
• 32:49
  verticalized agent, it doesn't. It doesn't do anything.
• 32:51
  Yeah, our vision with Sophia, it's like we want to be the
• 32:58
  channel of choice of our customers.
• 33:00
  We want our customers to like independent from our platform.
• 33:04
  They can do all they want to do in Sophia.
• 33:07
  So that's why we want to cover all the, the, the travel cycle
• 33:13
  and Sophia has each agent or has many agent for each use case or
• 33:19
  for each part because our our goal or our dream, it's to be
• 33:23
  the channel of choice of our customers like they want to talk
• 33:26
  with Sophia and that's it, it's. Like WeChat in China.
• 33:30
  Yeah, you. Can do anything in there and I
• 33:35
  want to break down to you got so you built Sophia, which is the
• 33:42
  travel assistant agent and you created it before there was Lang
• 33:49
  chain and before there was MCP. And so you made a lot of really
• 33:52
  interesting design decisions one being that you're going to
• 33:55
  create your own orchestration layer yeah.
• 33:57
  The other being that you basically created your own MCP.
• 34:01
  Yeah, at at the beginning we didn't have the orchestration
• 34:05
  layer because we don't have as many tools or that we, we have
• 34:10
  like the tools listed and the the LLM decide what to to call.
• 34:16
  But when we, we were growing more and more and more, we
• 34:20
  realized that we have a we we need a new like the
• 34:23
  orchestration and then multi or we need like a multi agent, the
• 34:27
  orchestration and then the the agents with their tools.
• 34:31
  So it was like kind of an evolution.
• 34:34
  We started with Sofia that calling directly to tools, then
• 34:37
  we evolved like creating Chappie and creating the satellite, the
• 34:41
  satellite agents with their tools.
• 34:47
  So it was like a decision made at the moment that when we
• 34:51
  start, when we were adding more and more tools, we realized that
• 34:56
  Sofia was getting a little bit lost there and we needed to
• 35:00
  evolve in some way. And when we did, we did the
• 35:03
  research and we we said, hey, we need an orchestrator here and
• 35:07
  there was any third party like very established in the market
• 35:11
  to to go. Because it was in like 2024 if I
• 35:14
  remember correctly. So yes, it was in 2024 we
• 35:16
  started the POG at the the POC at the the end of 2023 and in
• 35:22
  2024 it was like the first year with a full dedicated team to
• 35:27
  Sofia. And then you said now we need to
• 35:34
  recreate MCP or which at that time again there was no MCP, so.
• 35:39
  At the beginning when we when we create the orchestration layer,
• 35:43
  all the Asians were inside Chappie that it's our main app
• 35:47
  or our brain. But then we realized that
• 35:51
  Chappie was getting bigger and bigger and bigger and bigger and
• 35:55
  it's, it wasn't going to be able to manage that.
• 35:58
  So we decide to put away or like to take away the, the agents or
• 36:04
  that to the agents with their tools away from Chappie.
• 36:08
  And that's where when we decide like, we have to like create
• 36:12
  connections between these two things.
• 36:14
  So MCP were available at that time because it was in 2024 and
• 36:19
  and we grain like our own MCP. And now I imagine there's a lot
• 36:25
  of conversations between, do we upgrade to MCP or do we stay
• 36:31
  with what we've created because MCP has this ecosystem around
• 36:34
  it, or do we, and probably on the orchestration layer, you're
• 36:38
  probably thinking, do we even need an orchestration layer
• 36:41
  anymore? Can we just let the model cook?
• 36:45
  That's, that's, yeah, that's stocks that we that we have like
• 36:48
  every week. Let's say we now have MCP
• 36:51
  protocols in Sofia. We are not thinking about
• 36:55
  changing the old ones to MCPS. But as the world is more MCP
• 37:01
  driven now we, we have to or we had to add MCP to Sofia so we
• 37:06
  can keep up to date and connect with new things like it's a with
• 37:11
  external things or with our, with other APIs of our of desper
• 37:16
  guard that are creating their MCPS as well.
• 37:19
  So the new connections can be made from through MCPS.
• 37:27
  And about the orchestration layer, we are like when we, when
• 37:34
  you create an agent, like last week happens to us, we add a new
• 37:38
  tool with one agent like we create, we create a new tool and
• 37:43
  with examples of the use cases of that tool.
• 37:47
  We were to like, we weren't very specific and start with the
• 37:52
  Sophia the day next to the, the, the develop start to like
• 37:56
  sending all the prompts to that tool and that that wasn't right.
• 38:00
  So we think we need the orchestration layer for now
• 38:04
  because also the agents that we have are very different like
• 38:09
  issues assistant or flights or trip planner or dream facing
• 38:13
  like they are very different use cases that you have.
• 38:18
  So the orchestration layer keep that very optimized and very
• 38:24
  let's say simple in a way like breaking the decision of the LLM
• 38:29
  in two parts, keep the the the flow or the conversation more.
• 38:35
  Yeah, it seems like it's a separation of concerns.
• 38:37
  Yeah, yeah, it'll. Allow it to almost be like a
• 38:44
  walled garden, maybe not walled garden, but just if it's going
• 38:47
  to grab one of these tools, then it takes it.
• 38:51
  I still feel like you probably if you reassess in a year or
• 38:55
  maybe six months, it'll be like, well, OK, maybe we can get rid
• 38:59
  of it now. Yeah, but considering how much
• 39:03
  like it's just wild I how much you don't need them anymore for
• 39:09
  a lot of the use cases. Right.
• 39:10
  Yeah, Yeah, that's true. But as you mentioned, we saw
• 39:14
  Sofia like 2 years ago in a desire to combine our knowledge
• 39:19
  of assisted contacts in, in sales, in after sales.
• 39:23
  Like we saw the opportunity right there and it was so clear
• 39:27
  like LLMS were the the perfect use case of LLM were
• 39:31
  conversational. We have been selling tickets or
• 39:35
  flights conversationally. So it was like the perfect
• 39:39
  combination of those two words with the know how that despair
• 39:42
  has on how to do that with human agents.
• 39:47
  Like we are very in touch with them and they also make my my
• 39:53
  job easier. Like saying, Hey, we thought
• 39:56
  that like a human nation would do this in, in this interaction.
• 39:59
  And we took it. We saw if that use case, it's
• 40:03
  also valid for Sophia and we implemented.
• 40:06
  So it was like a desire to combine those two words that are
• 40:11
  very deep in our culture. Do you ever give Sophia the
• 40:15
  ability to bargain with the customer or negotiate?
• 40:20
  Because I know that on certain hotel booking sites they'll say
• 40:26
  we match the lowest price. Not yet.
• 40:29
  We didn't do that yet. Because it could be dangerous,
• 40:32
  right? It could be dangerous, Yeah, you
• 40:33
  have the LLM is is deciding for you, so could be dangerous.
• 40:38
  Yeah, and you think like somebody can just come and say,
• 40:42
  oh, I saw it cheaper on this website and the LLM is like, OK,
• 40:46
  I'll give it to you for $1.00. Yeah, yeah, yeah.
• 40:49
  So no, no, not yet. We are not thinking about that.
• 40:53
  Yeah, it'll be nice when it does happen.
• 40:56
  Tell me so. I.
• 40:57
  Can I would remember this conversation?