Building Phone Call Agents | Course Introduction

Okay. So, hello everyone. We are finally live in this listen lesson zero live session where I'm joining my dear friend

Jesus Copo. And today we are just going

Jesus Copo. And today we are just going to introduce this uh discourse which I think you are going to to love it. Phone

calling agents.

So I'm going to give you're going to give like a couple of minutes for >> So, let's, double, check, that, that everything is working. So the streaming

>> all, the, streaming, the, audio, is, fine, the image is fine video is fine and all that stuff >> I, think, your, YouTube, is, fine, ML, I, think I'm I think that's working. Let me just

check stack.

>> Yeah, let's, also, check, LinkedIn., I, think Substack is still not connected or at least not streaming.

>> Okay,, YouTube, seems, to, be, working.

>> Yes,, YouTube, for, sure., Yeah,, I, can, see that.

>> Okay, let's wait for a bit more comments to see that everything is working fine >> to, get, all

>> Okay,, I, can, see, your, comment, here., teams

working give us.

Yeah, just texting here in YouTube just to tell the people to wait a minute.

>> Cool.

>> Let's, see, if, it, link, it, in., Let, me, just double check that as well.

>> Okay., Yeah,, seems, to, be, a, problem, with with Substack.

unable to connect.

>> Yeah,, LinkedIn, is, working,, right?

>> Yeah,, I, think, LinkedIn, is, working, and also uh and also YouTube. But let's try out uh Substack another time because there seems to be a problem here.

>> Yeah.

>> Okay.

>> Maybe, just, drop, a, drop, a, message, in Substack with the link to you to go to linkering >> and, then, we, will, fix, it, in, the, next, one for sure. I think we are sending data

for sure. I think we are sending data right now to Substack. But yeah, it seems to be giving some problems. So yeah, I'm just going to tell the people

from Substack to join YouTube.

Strange. We tested this. [laughter]

But it's all right. I mean, there are other platforms. They can join from other platforms. It's completely fine.

Okay let's Okay, so I'm just going to send the YouTube audio, the YouTube video on

Substack and we are beginning. So sorry

for the delay.

Okay, here we go.

Perfect. So

>> okay., [laughter]

a bit of delay but we are ready to go now. Uh these are the things related to

now. Uh these are the things related to to streaming. It's our first time so

to streaming. It's our first time so [laughter] sorry for that. Sorry for

that. But but yeah, uh so yeah, I just we just prepared here uh a couple of slides that I think

you're going to to enjoy where we are just going to to show you the top content of the lesson zero and also what you can expect out of this

course. Remember that this is just

course. Remember that this is just lesson zero where we are outlining the different lessons that we are going to cover. In total there are going to be

cover. In total there are going to be four sess four uh lessons and also a couple of short demos that we are going to cover just to give you uh

excitement about what is to come but also we are hiding some stuff because we don't want to spoil the surprise as the as the course advances. So sorry for

that guys but yeah that's what happens.

Uh so yeah we are going to start with these uh slides. So, Jesus, I think I'm going to hide us both from here because we are

>> okay., But, then, we, will, appear, again.

>> okay., But, then, we, will, appear, again.

So, yeah, first of all, even if you already know him uh who doesn't know uh Jesus Copo, I just wanted to yeah to introduce to formally introduce Jesus

Copo.

I think Jesus Copo combines two very important stuff. uh first of all he has

important stuff. uh first of all he has real experience he has been working in the AI field for a lot of years and also he combines this with uh high quality in

terms of teaching advanced concepts that's the reason why uh I think our philosophy aligns pretty well and that's the reason why he is part of this course and he was also part of faba so yeah

Jesus I don't know if you want to to add something to this woman introduction but go ahead >> no, it, was, amazing, thank, you, thank, you, m for the introduction I just want to say that I'm super excited to be here with

all of you guys. I think this course is going to be amazing and yeah, just about you Miguel, I'm really happy to to be collaborating with you again. I really

had a a lot of fun building the project and just building all the content around Ava and I think this course is going to be even bigger and even better. So super

excited.

>> I, agree., I, agree., I, think, it's, going, to be to be massive. And we we had to be honest, we had a lot of a lot of fun building this to be honest. So yeah.

Okay. So as as we were commenting uh Jesus and I in case you are new to the neural maze and to our content. So we

built this course before Ava the WhatsApp agent by the way I have the like a list in Substack where you can go

through all the lessons. This was a a pretty, long, course, to, be honest, Jesus because it was about six lessons. This

course is going to be a bit shorter, but yeah, if you want to know our previous work, you just need to to put here the >> Yes.

>> Yeah,, the, all, the, stuff., So,, yeah,, I'm just going to add here our faces. Okay.

So, let's keep going with the with the slides. And now, let's get into the real stuff. Let's get into the interesting part. So, as I was saying

interesting part. So, as I was saying Jesus and I built in in the past this ABA, the WhatsApp agent, which was really cool. I mean, we really had a

really cool. I mean, we really had a great time connecting agents with WhatsApp. But then we had a thought. We

WhatsApp. But then we had a thought. We

thought it would be really cool to bring this idea one step further and just try to create a real time communication.

Okay. So this is the result of our project. Basically a complete agent

project. Basically a complete agent powered call center for a real estate company. Okay. It's going to handle both

company. Okay. It's going to handle both inbound and also outbound calls seamlessly through Twilio and also fast RTC. Agents can search the property

RTC. Agents can search the property database in real time with filters across numeric string and categorical data thanks to superl. and it uses open

source TTS and SD models for full speech integration. We will also show you how

integration. We will also show you how to do this, how to deploy your own TTS models and SD models to Rampot to the GPO cloud and yeah, we will also show

you how to use this Orphus 3 billion that we detected and we verified that it had uh amazing outcomes and if you check >> sounds, really, good.

>> Just, wanted, [clears throat], to, give, a short introduction to the to the diagram in here. So let me get the pencil. So

in here. So let me get the pencil. So

the first thing if we follow the numbers number one it's our project repository and in here let's take out our faces for

this one. And in here in this first step

this one. And in here in this first step we are just going to show you how to deploy the both like two docker images. One for

the API that we are going to build.

Okay. and the other one for the image that is going to be injected that is going to be used from ramport. All

right, the first one here this one is a fast API application that has web RTC and websockets connections and it's going to be used by fast RTC. That's

something that we are going to cover in lesson one and we'll give you some some introduction today. And the other one is

introduction today. And the other one is this image in here which combines llama cpb right with the orus 3 billion guf

all right that's basically what we are going to build if we follow this uh of course we implemented devops with github actions and we are deploying both docker

images to the docker hub of course you can replace this with ecrm amazon uh the good thing of the code that Jesus and I have been building is that you can

adapt this to any cloud provider that you want. You can uh modify this and

you want. You can uh modify this and adapt it to any use case. So that's uh that's something really positive we believe.

>> Exactly., Yeah., Completely, agnostic.

>> That's, it., That's, it., And, and, yeah,, in two we are deploying both here in Rampod and also here. There is not a line for this one because it was a bit messy in

the diagram.

But yeah, the first image is going to be fetched in rampot inside rampot by a llama cpp server. All right. And in here we are going to expose this v1

completions. This is the orus 3 billion

completions. This is the orus 3 billion tts model for the models that we are going to show you today and in lesson one next Wednesday. You can see that

everything will be able to run in your local laptops. for Oruse 3 billion

local laptops. for Oruse 3 billion that's not possible and that's the reason why we need to uh rely on cloud providers like ramport but of course as

we mentioned as I mentioned you can also uh do it with any other GPU provider or even Amazon cloud assur whatever you

want so yeah this is uh number two so now let's go to the left part of the diagram in here you can see uh like

three guys uh starting uh phone and this This is the place where Twilio will become really important because the phone call is going to be sent to a

Twilio number. You don't need to waste

Twilio number. You don't need to waste any kind of money in this uh because I mean if you haven't used Twitter before you will have a free number with your free account. So you can use it whatever

free account. So you can use it whatever you want. In case you have exhausted

you want. In case you have exhausted this free account, then probably you will need to to purchase your your number. But yeah, the phone number

number. But yeah, the phone number arrives to the Twilio application and then it starts sending post request to this

endpoint that we are enabled in the fast API this voice telephone incoming.

Okay. And all of this is going to receive 2ML instructions. That's

something we are going to cover throughout the course. that these 2ML instructions are the way in which the 2ML applications the applications that you build inside Twilio are going to communicate with our back end. All

right. After that, this is something that we have covered a lot of times.

Jesus in his videos and and my articles.

We are just passing the audio the audio frames through the ST and then we are just uh computing the response.

Let's put our faces now. Computing the

response here in the fast RTC agent.

This is uh really cool because we have combined fast RTC with langraph. uh we

believe that it's a very cool combination because you can uh take all the advantages of langraph as a pretty robust framework and also fast RTC for the real-time communication. Fast RTC as

we will see in lesson one it's basically a way for us to uh avoid all the pain involved in dealing with websockets RTC etc. It's like a layer, an abstraction

layer. And

layer. And >> exactly.

>> Jesus, is, a, big, expert, in, this.

[laughter] >> Well,, uh >> I, tried, I've, done, a, couple, of, things >> and, yeah,, after, the, ST, the, transcription arrives to the React agent to this fast

RTC agent, the response is going to be sent back with this TTS module in here.

All right. And this TTS is just providing the audio response with a personalized voice. All right. And

personalized voice. All right. And

that's basically what we are going to build throughout this course. It feels

like a lot. Well, uh two things that I forgot. Of course, we are going to add

forgot. Of course, we are going to add the capability with my beloved OPIC both from versioning traces evaluation data

sets for the evaluation etc. And also and this is a very cool stuff. We are

going to add a tool to our agent that is going to provide real time search. In

our case, remember that we are building an agent power call center for a real state company. So we need properties.

state company. So we need properties.

Properties we can search in a vector database. And the reason we chose super

database. And the reason we chose super linked is that if we solely relied on quadrant, if we search for things like can you

give me one property in Bario Salamanca in Madrid less than €300,000 with three rooms, two bathrooms, and two

living, rooms,, for example.

These kind of really specific numeric searches are going to be solved pretty accurately with superlink. It's just

like a layer of substraction on top of quadrant that is going to provide the agent with tools to search the property that we want in a very efficient and exact manner. All right. But by the way

exact manner. All right. But by the way I didn't uh mention it, but all the questions that you have, drop it in the channel in the in the comments because

by the end we will read all the relevant uh questions and answer any question that you have >> and, we, leave, like, 10, minutes, in, the, end.

So no >> 10, minutes, uh, more, or, less., Yeah., All

right. So this is in a nutshell the application that we are going to build.

All right., So, now, let's, keep, evolving.

because now I just wanted to uh to mention the structure of this course. All right. So this course is

course. All right. So this course is built around three fundamental resources, right? It has a GitHub

resources, right? It has a GitHub repository.

It has one article per lesson and it has one live session per lesson.

The live session is going to be pretty similar to the one we are having uh right now. But this is just an

right now. But this is just an introductory live session. The rest of the live sessions are going to be more like walkthroughs uh similar to the content that Jesus and I are putting on

YouTube and Substack on a weekly basis.

Okay, perfect. So

about the GitHub repository just wanted to share you here the GitHub repository because it's already available. You can

look for it under neural maze organization real time phone agents course., All right., And, in, here, we, will

course., All right., And, in, here, we, will push on a weekly basis a release.

>> Exactly.

>> All right., So, here, you, can, see, the, first release this week zero project overview that contains basically just uh links to

the substack article links to the live session that we're having today and basically that all right and in here we will be posting as time

goes by in this course breakdown week by week we will be posting all the relevant information all the links that you need okay to to interact with the code and to

interact with the resources that we are about to to build. So to summarize, next Wednesday you'll have a new release in here week one with all the relevant

content about fast RTC langraph agents.

All right,, we, will, push, all, the, code related to how to connect fast RTC with langraph and with the with the agents basically with these property agents.

Okay, so let's go back share this tab instead. So yeah, let's go back to to this slide. So as I was saying, every Wednesday we publish a new

release. The code is going to be fully

release. The code is going to be fully aligned of course with the written lesson >> and, our, idea.

>> Exactly., This, is, good., Also, just, got, we just got a message in the in linking just asking about the course. So is the course like a live stream live stream based? No. So the live stream is just

based? No. So the live stream is just >> something, else,, right?, So, the, so, you're about to say anything else. of the

course is mainly written articles in SAS.

>> Yeah,, that's, it., That's, it., That's, a very good question. Uh it's Robert right? Yeah, Robert Dean.

right? Yeah, Robert Dean.

>> Yes.

>> Yeah., Yeah., Uh, I, think, that's, a, a, very good question. And the thing, Robert, is

good question. And the thing, Robert, is that we have noticed that people have like different preferences. Some people

prefer written articles, some people prefer these live sessions, some people even prefer to go directly to the code because they are pretty advanced.

[laughter] Yeah.

>> So, that's, the, reason, why, we, decided, to go with these three formats. Okay. And

>> we, give, it, all.

>> Yeah., Yeah., We, give, it, all., We, give, it all. And yeah, every Wednesday, as I was

all. And yeah, every Wednesday, as I was saying, you'll have the written article and the code in case you want to ignore the written articles and the live sessions. You can go directly to the

sessions. You can go directly to the code.

On Sunday, we will have a live session about the code and about the article that we have uh put on Wednesday. And of

course, and this is important, the code is going to be 100% open source. So if

you feel like uh really proficient really advanced and you don't need explanations about WebRTC or websocket agents, etc., you can go simply to the to the code and just use it for your own

purposes. Okay, I can see

purposes. Okay, I can see >> and, if, you, build, something, cool, with this as a base, just let us know, tag us. We want to see we want I think you

us. We want to see we want I think you can build super cool things with this base. So once you understand everything

base. So once you understand everything this is really cool. You'll see.

>> Okay., So, I, think, people, were, asking, for the for the reple. So yeah, here they are.

>> Yes,, the, link.

>> Yeah,, if, you, want., Uh, you, can, share, it again. Uh Jesus.

again. Uh Jesus.

>> Yep., I, can, share, that, in.

>> Okay., This, is, a, a, good, question.

>> Are, you, going, to, write, and, run, this, on Google Collab? So uh yeah there are

Google Collab? So uh yeah there are going to be different parts of resources inside the the lessons. For example for next lesson on Wednesday we are going to

split things. We are going to teach you

split things. We are going to teach you the basics of fast RTC using Google Collab Jupyter notebooks etc. And then once that's clear we are going to move to the code to teach you how to move

from Jupyter notebook to to basically to the code to build something more production ready.

Can we use model in place of ramport? As

I was saying, this course is going to be focused on ramport. Having said that you can replace whatever piece of uh technology that you want. One thing that

Jesus and I uh are obsessed with is to make things modular, to make things uh I mean we all know how this feel evolves.

So one day one thing is super hype and the the next day is completely obsolete.

So So yeah. Yeah, absolutely. You can go with with model and in fact if you want and if you build something with model we will be happy to share it in our social media.

>> Yeah,, even, just, create, a, PR, and, maybe, we can even have absolutely sure >> we, got, we, got, some, PRs, in, ABA, as, well.

[laughter] >> Absolutely., And, one, of, our, sub

>> Absolutely., And, one, of, our, sub subscribers even um translated all the ABA project into Amazon like 100% Amazon based. So yeah uh we are happy to to

based. So yeah uh we are happy to to share the things that you build.

Absolutely.

Okay. So, I think there are a lot of questions. So, maybe

to to note down all of them and just >> Yeah,, let, I, I'll, I'll, be, doing, that., So,

let me just do that.

>> Okay.

>> Because, we, also, got, this, interesting question. So, I I'll I'll take care of

question. So, I I'll I'll take care of that., Let me, just, continue, with, this

that., Let me, just, continue, with, this >> because, we, have, an, hour.

>> Yeah., Okay., So,, so, yeah,, we, are, going, to go really fast uh onto this substack articles. As I mentioned, every

articles. As I mentioned, every Wednesday you will have an SAP stack article. The only difference with the

article. The only difference with the code is that Substack articles are going to be really detailed with a lot of resources, a lot of work under them. Uh

so these are going to be for premium subscribers of the new maze. All right

live sessions more the same. This is the first session of all but after that we are going to uh to isolate or to reduce these live sessions to the premium

subscribers and one of the reasons is that we believe the experience is going to be better uh because if we open this to a lot of people a lot of questions a lot of noise is going to to be

introduced so so yeah we wanted to make this uh for the premium subscribers but as I mentioned before the the code is 100% ready for you to to take it. I mean

we wanted to to keep the open source vibe.

Okay. Uh I can see here that people are jo are asking how to join this course how we can join you mains WhatsApp and

any kind of community. Uh so what I'm going to do is that after this uh after this stream is ready we are going to

send or by the end of this stream we are going to send the link to the to the new maze. Okay. to the Well, in fact, I can

maze. Okay. to the Well, in fact, I can do it right now.

>> Yeah,, I, can., I, was, about >> Oh,, yeah., Yeah., Yeah., Go, ahead., Go

ahead. You can do it.

>> Let, me, just, put, lesson, zero, here.

>> Yeah,, that's, it.

>> YouTube and also in LinkedIn.

All right., So,, that's, lesson, zero., So,

that's the lesson that is available to everybody. And then if you want to um

everybody. And then if you want to um get the the next lessons, you got to be a subscriber.

But open source uh code that is for everybody as well.

>> That's, it., Uh, and, yeah,, Praep, Kumar,, I can see that you're asking how to join as premium subscribers. If you click the article that Jesus cop has just uh shared, you will see all the details in

there. How to uh how to join as premium

there. How to uh how to join as premium subscriber, to, a, newsletter., All right.

So, >> lastly,, uh, we, are, just, getting, into, the calendar. All right. This is going to be

calendar. All right. This is going to be a pretty long course uh because it's going to be spanned throughout four weeks. The first lesson is going to be

weeks. The first lesson is going to be about building fast RTC langraph agents and the article is going to be received on November 19 and the live session next

Sunday. Lesson two is going to be about

Sunday. Lesson two is going to be about superl and how to build uh this realtime property search using superl. All right

this is going to be on November 26 and November 30 for the live session. On

lesson three, and this is our favorite lessons probably, we are improving the ST and the TTS systems using Orfuse 3 billion, teaching you how to deploy this

into Rampot using Lama CPP server and also teaching you how to build custom models that can be attached as plugins to fast RTC. This is probably one of our

favorite uh pieces of code. uh the

article December 3 and live session December 7 and finally the final deployment. We will show you how to

deployment. We will show you how to deploy everything to Rampot both the API and also the the model monitoring all the

observability layer that I was talking to you about previously with with OPIC and finally the the final two integration where we will show you how to create the 2ML app and connect it

with the with the API. Right.

>> And, just, a, quick, note,, Miguel,, the, live session, so all the live session will be recorded and will be sent to you. So you

don't really need to be at 5:00 pm every Sunday to to receive the live session.

So of course, if you want to ask questions, if you want to build with us if you want to if you're building this project and you want to have a space to have a to ask a question, then yeah that would be the place to be. But

otherwise, no worries. The session will be recorded at 17.

>> Absolutely., watch, any, time, and, and, you can keep that forever.

>> Very, good., Very, good, remark., Uh, Jesus.

Yeah. Yeah. Absolutely. You will receive an email with the recorder session if you are subscribed to an M. Yeah.

Absolutely.

All right. So for this lesson zero, we weren't sure on what to to discuss. So

we decided to give you like a preview of each one of the lessons. So on lesson one, fast RTC agents, we are going to build something like this. Okay. Next

Wednesday if you follow up all the article and you join the the live session you will build something like this. We name it the fast RTC agent and

this. We name it the fast RTC agent and if you check inside is just a react agent of course as any react agent with a set of available tools but the

difference is that it has attached some string handlers. All right, these stream

string handlers. All right, these stream handlers, I'm not going to get into detail because that's something that you will you will see next next week, but these stream handlers are the way in

which fast RTC enables real-time communication. Okay? And you can use

communication. Okay? And you can use already built-in stream handlers or as Jesus did for this course, you can build your, own, stream, handlers., All right?

>> And, we, will, teach, you, how., Yeah.

>> And, I, will, teach, you, we, will, teach, you how. Uh also a short-term memory as

how. Uh also a short-term memory as always. And the input in this case is

always. And the input in this case is going to be audio frames that are going to be translated. They're going to be transcribed. And finally, we get as

transcribed. And finally, we get as result this uh I don't know if you're seeing this uh this DTS. Okay. One of

the things that I wanted that we wanted to teach you in this first lesson is the theory behind

fast RTC of course, but also about VAT.

Okay, voice activity detection. All

right, because this is one of the basic tasks that you need to understand to really understand how real communication with phone calling agents uh works.

Okay, >> turn, detection, and, also >> yeah, that's, it, and, in, here, you, can, see basically the like the workflow that we are going to follow. By the way, we are using this

follow. By the way, we are using this fantastic uh GitHub repo siler robot uh because this is the one that is going to be used by fast rdtc and we really like this uh this repo. So it was really cool

when we discovered that fast RDC was using that under the hood and yeah basically we receive speech a speech input like audio frames and then the

algorithm the bad algorithm is going to detect when we start talking and when we stop talking all the audio frames are going to get accumulated and they are going to get sent to the agent. This is

a capital part of the system that we are about to build, right? Because we need to detect when the human the human starts talking and stops talking. And

that's exactly the audio that is going to be sent to the agent to the ST and then, to, the, agent., All right.

Yeah. Thanks, Jesus, for the >> just, with, a, pin., [laughter]

>> Yeah., uh, more, on, the, ST, and, TTS, models that we are about to build because uh we were talking in a very abstract way

about these two models but for the STD in this first version we going to use moonshine models then we will update

into the whisper family of models for the DTS we are using Kokoro which is a pretty small model in terms of today's standards I think it's 82 million Jassus

more or less I don't remember >> I, think, exactly, 82, million, yeah, so, coor is 82 million >> 82, million, parameters, so, it's, pretty small but it's really really efficient and the results are pretty cool if you

take into consideration that you're running everything locally so that's what we are going to build in next

week's lesson right oops and yeah by the end of next lesson by the end of lesson one you will have something like this running on your computer. Okay, this

radio UI that lets you interact with your agent, right, with your voice agent. In fact, I have here, I mean

agent. In fact, I have here, I mean this is a bit of a spoiler, Jesus, but I have here uh the lesson, the notebook that you are going to receive next

Wednesday. And I can see I can show you

Wednesday. And I can see I can show you a little bit about the the things that we are going to cover. Um, and I think it's interesting because for those of you who are not enrolled into this course, maybe you can see the kind of of

of content and the kind of detail that we want to get with our courses. But

basically, this first notebook is going to be about fast RTC the core concepts, the different streaming modes, modalities, the different type of handlers, streaming

handlers, as sync handlers. Uh, and then we are building from the very basic handlers like this echo audio. This is

just going to uh echo my audio. Just let

me know, Jesus, if I'm sharing the the audio right now. So, let me see. See?

Okay. Hello. Okay. Hello.

>> Can, you, hear, me?, Can, you, hear, me?

>> Yes.

>> Okay., So,, this, is, going, to, be, like, the first example. I'm going to take the

first example. I'm going to take the other frames and I'm going to return the other frames back. And then we are going to keep evolving, keep making things more and more complex. You're going to

start using the VAD, the voice activity detection with the reply on post handler. We are going to start adding

handler. We are going to start adding TTS and ST like for example this is a funny one that Jesus and I uh came up with uh which is just going to copy what

you say but with a coor voice. All

right. So let me check here. Uh Jetty

hello how are you?

>> Hello., How, are, you?

>> You, are, copying, my, voice.

How are you?

>> How, are, you?

>> Okay,, so, you, can, see, how, this, works.

Basically, in this case, we are just taking the my voice. We are transcribing my voice and then we are taking this text and passing it through cocoo to generate a new audio to generate the

voice that we are receiving. All right.

And yeah, and then but I didn't want to show you this because it will be a spoiler, but we are adding LLM calls. We

are adding agents but without tools. And

finally, we are adding a react agent tools and something that Jesus came up with which is really cool that when the agent is using a tool, we are like

breaking the loop and like interacting so that it feels like the agent is actually searching for something. So we

can put whatever audio you want like hey uh let me take a look at the system to find your property. Something like that.

We we thought it was pretty pretty funny and we decided to put it Uh look this is a very very good question that we need to to take Jesus like for the project.

>> Mhm.

>> Uh, you, taking, into, account, only, English voices or can we use other languages like Spanish using the same tooling you chose. So right now we are focusing on

chose. So right now we are focusing on English but as mentioned before you can replace the TTS models, you can replace the ST models, you can replace whatever

you want using even your fine-tune models for another kind of language and potentially >> you, can, even, fine-tune, a, fine, tune, a, TTS on your own voice.

>> Yeah.

>> And, then, so, that's, something, that, you can you can even do. Maybe we create a bonus lesson that would be outside of the course. But yeah, um but answering

the course. But yeah, um but answering this question for sure, um the ST whisper is multilingual, so that's for sure. And then the TTS model, as Miguel

sure. And then the TTS model, as Miguel is saying, sorry for interrupting Miguel.

>> No,, no,, no., Go, ahead., Go, ahead.

>> Some, some, some, TTS, models, are, capable, of um Spanish as well, and some others are not, but we will show you everything.

>> Absolutely.

>> Another, one,, another, question, that, I like uh can we add WhatsApp video call something like that in future? So that's

a a pretty pretty good one because fast RTC if you go to the beginning you will see that we have audio audio video and video. So

video. So I think that's a potential idea. I mean

I like the idea to be honest.

>> I, like, the, idea., Yeah.

>> Yeah., Sounds, like, an, interesting >> it, can, handle, it, can, handle, video stream. So you can I don't know just

stream. So you can I don't know just have a system that takes a screenshot every 60 seconds or something like on a voice command. It can take a screenshot

voice command. It can take a screenshot and then just send that image to the model. You can build cool things. You

model. You can build cool things. You

can really That's a really good question. [laughter]

question. [laughter] >> Absolutely., Absolutely., We, love, your

>> Absolutely., Absolutely., We, love, your ideas, guys. Yeah. Yeah.

ideas, guys. Yeah. Yeah.

>> Yeah., Yeah., You, you, give, us, ideas, for new content. So, we really appreciate it

new content. So, we really appreciate it [laughter] to be honest. Sometimes it's

hard to come up with these projects.

>> Okay., So,, so, yeah,, let, me, go, to, the slides again. Well, yeah. I think just

slides again. Well, yeah. I think just before we go to to Super Link, I just wanted to share you with you the the kind of application that you will have

uh by the end of next week. Okay. So

let's hide this comment. This is the kind of grad UI that we are going to cover uh next week. Okay. By the end you'll have something like this running on your computer and you will be able to

interact with it.

>> Well,, let, me, Yeah., Yeti, stereo.

>> Welcome, to, the, neural, maze, call, center.

This is Tara, your favorite voice agent.

>> Hello,, Tara., Nice, to, meet, you., I'm

>> nice, to, meet, you,, Miguel., I've, always wanted to be an AI engineer. The neural

maze feels perfect.

>> Okay., So,, so, yeah,, I, think, Cara, is, way too kind, but uh but yeah, you can see the kind of interaction that you will have. Uh it's pretty pretty uh I mean

have. Uh it's pretty pretty uh I mean the the latency is small, but the thing is that I'm running everything locally in the future lessons. We will deploy things to the GPU, we will deploy things

to to the cloud. So everything will be uh will be faster. Okay. And the voice >> even, better, even, even, the, voice, would, be even better.

>> This, is, Koko, and, the, ST, is, moonshine., So

we are with the worst model that we are going to use up. I mean in future lessons everything is going to be better., I, mean, Orus, 3 tribillion, is

better., I, mean, Orus, 3 tribillion, is going to be pretty amazing but we didn't want to spoil the surprise today to be honest. Uh so maybe in next live

honest. Uh so maybe in next live streaming we will give you a hint of how the voice is going to to look like. Uh

but yeah, this is what we're going to to have in in next lesson.

Okay. So for lesson two we are going to get into super linked.

All right. We are going to understand why superlink is necessary for the property search. And we are also going

property search. And we are also going to understand how super link connects with pretty famous uh vector database that I have used and Jesus had used in

in many of of our projects quadrant.

And the thing is that we are going to be able to search for a lot of different data types, a lot of different data associated with the properties in uh

very uh in a very efficient and fast queries. Okay. So let me show you now

queries. Okay. So let me show you now the kind of thing that we will have by lesson two. But your quadrant

lesson two. But your quadrant uh I mean probably you are familiar with uh with quadrant with the dashboard but basically here you can access your different

collections and see the vectors that you have, stored, inside., [clears throat] So here you can see the kind of elements the kind of entities that are going to

populate your quadrant collection after you add the super linked on top. Okay

here you can see all the different properties that define our real estate properties. Okay, first of all, we have

properties. Okay, first of all, we have this description. This is of course

this description. This is of course text. This is going to be embedded as

text. This is going to be embedded as normal text. But then we have another

normal text. But then we have another property which is rooms size in square meters, the city that this property belongs to, the district

the price in euros, uh images because this is something that Jesus will show you in some bonus lessons, but we also provided some multimodal capacity. Uh

also the schema, the object ID, etc. And we have a lot of different properties.

Sorry guys because we focused on Madrid because we are from Spain but uh of course you can do the same with any any country that you want. Uh but yeah we Oh

look I used to live here uh when I was living in Madrid. Uh such good memories good memories. Uh and yeah this is

good memories. Uh and yeah this is basically what you will what we will get in quadrant. Then if we go back to to

in quadrant. Then if we go back to to the slides let me share the tab.

This super linked everything that I've been sharing with you, all these property index, the property queries that we are going to define are going to be connected with our agent with the tool. Okay? So whenever the agent is

tool. Okay? So whenever the agent is going to search for a property when we ask about I don't know give me a beautiful apartment in Madrid the agent

is going to search in superl and it's going to add some uh magic that says hey let me look for this in the in the in

the system for example or you you can tune the kind of audio that is going to be displaying there and we wanted to do this because if you have tried systems like this one in the past. I

mean, I've tried a couple of these applications that try to do the same as we are doing. It's really weird when you ask for things in a database and the

agent says nothing. It's like silence. I

mean, when you're talking with a human being, at the end of the day, the human being is going to say something like "Okay, let me check." Or, "Uh, can you give me one second, please?" So, we wanted to replicate the same kind of

behavior in our system. All right.

Oh, I I really like one of the questions, Jesus. Uh, let me

questions, Jesus. Uh, let me >> Okay,, which, one?

>> Put, it, in, the Okay, so it's in the place of super linked, can we use knowledge graph? So

that's a very good question and that's something that we thought at the beginning like should we go with superl or should we go with a knowledge graph?

Of course, we chose uh superl, but to be honest, I think it's uh possible to do it with a with a knowledge graph. And in

fact there are some things that probably are better covered with a knowledge graph approach all the entities all the relationships. So yeah I think that's a

relationships. So yeah I think that's a yeah that's a good good idea and if you manage to to do it with a knowledge graph like Neo forj or something like that it will be it will be cool uh to check it out.

>> Exactly., I, think, it, really, depends, on the project that you're building upon or the system or the industry. Right? In

our case is going to be this demo is going to be about properties. And here a cool thing about Superink is that I think Miguel explained that you don't only you are not only um how do you say like creating the embedding for the text

description. You're creating the

description. You're creating the embedding that is combining the the text description and also all this numeric data. So when you ask for can you give

data. So when you ask for can you give me a property similar to this it's not only going to take the description for the semantic embeddings right for the for the semantic similarity it's also going to take those numeric things so

it's going to give you properties that more or less have the same rooms more or less have the same price and you can even weight everything it's super cool and >> that's, really, cool

>> that's, really, cool, okay, this, is, another interesting uh let me put here another interesting question so we are going to change uh Cocoo to Orphus which is a bigger model.

Uh that's exactly what we are going to build. We are going from 82 million to 3

build. We are going from 82 million to 3 billion parameters and and yeah uh you just need to check the the audio results

because I mean they are pretty amazing and I'm going to link with another question that is more or less related.

Does we add age personality to voice? So

not exactly like this but we have like uh I think seven eight voices to choose from male female uh like uh older younger etc. One of the things that

Jesus mentioned and we are thinking like really seriously for future lessons is to clone our own voices like show you

how to clone your your own TTS your own Orphus 3 billion on your own voice using Anslot, for example, generate, a, GUF, and create a deployment script that is able

to deploy the same thing to to Ramport.

So let us know if if you you will be interested in in this thing.

>> But, we, initially, take, it, out, from, the course because we thought it was probably too much uh to digest.

But I mean we are really into fine-tuning. We we would love to to

fine-tuning. We we would love to to teach fine-tuning for uh TTS models because that's something that I don't think a lot of people are doing right now. and and yeah, we were happy to to

now. and and yeah, we were happy to to add some bonus lessons about uh fine-tuning TTS with u different voices even cloning some uh famous voices. Uh I

don't know, just let us know what you think.

>> Let, me, just, quickly, answer, the, last question from Techlan and then just let's be also wary of the time is 15 minutes to six. But yeah, just this quick question because I think it's

important and I left that behind and and I think Techlan is just answering giving the question again. It's about why we chose fast RDC instead of like it and pyat which are two really important and

two really solid and robust frameworks for um real time communication. We chose

faster because it's lightweight, it's light and it's modular and we can inject anything that we want. For example

langraph. So you can inject any agent framework to do like the backend things right? And with um live and pyat, you

right? And with um live and pyat, you really got to use their voice agent framework which is not as flexible as as langraph and we just like the idea of having a different interface. So you

have the interface that is text and if you want you can have um a voice on top of it but still the agent and the back end is the same. So that's why we chose

it.

>> Okay., So, let's, keep, going.

>> Yeah., 15, minutes.

>> Yeah., Yeah., Yeah., Uh, okay., No, look, this is an an important one. So how does VA differ from automatic speech recognition? So uh you could understand

recognition? So uh you could understand automatic speech recognition is a much broader field and one of the techniques is this voice activity detection because there are other techniques there are

other tools to to do the same. So I

would say like ASR it's uh in fact uh I know I'm running late Jesus but forgive me. It's all right

me. It's all right >> here., I, left, this, uh, this, source,, okay,

>> here., I, left, this, uh, this, source,, okay, this link in here that contains like a huge it's like the bible for ASR

automatic speed recognition and uh etc. So, so yeah, uh John Michael, I would suggest, you, to, to, go, here,, all right,, to join the the Substack and automatically

you will receive all the free or premium. I mean this this video is going

premium. I mean this this video is going to be uh received by both uh free and premium.

You can check the slides and I think it's a pretty pretty interesting way to to relearn the difference. I mean it's like for me the difference between AI and deep learning or the AI and machine

learning like they are like uh subsets and superersets. So so yeah it's a bit a

and superersets. So so yeah it's a bit a bit different.

All right. So about this lesson three I need to uh spit uh a little bit. Uh so

from moonshine we are going to go in lesson three to the whisper family of models and from cocooto as I as we mentioned we are going to oruse 3

billion and we will deploy the orus revilion into ramport using llama cpp.

All right using by the way guf uh model format. In fact, I just wanted to show

format. In fact, I just wanted to show you here uh share this step instead.

Yeah. So after lesson three, you will have something like this on your rampod uh, dashboard., All right., You, will, have

uh, dashboard., All right., You, will, have this Oreus server that is going to accept HTTP well completions like the

OpenAI type uh API completions request from your uh fast RTC plug-in. Right?

It's a bit complex but with the code you will you will understand it. Basically

we are like sending all the different chunks of text and Orus server this llama CPP server is going to send back all the all the audio generated right and we are going to take all the audio

and send back this audio in chunks.

Okay. And that's exactly what you are going to to hear with Orus.

Uh yeah nothing much to add here. Uh by

the way in case you don't want to build everything from scratch on your local machine we are also going to share with you public docker images in case you

want to use for example this Oreus llama CPP server for different projects we are releasing these uh docker images because we we saw that many people are like

putting into private things like that and it's like we don't understand why uh so we are releasing uh publicly the the docker images for applications in case

you want to simply like use it for another project. I mean this is pretty

another project. I mean this is pretty pretty high performant to be honest. I

mean you can host everything on uh RTX like 5090 I think it was and it will cost you like 50 cents uh an hour or something like that in case you want to

put it on a pod. Ideally you will put it in some serverless application but you need to be careful with the latencies of course. Yes, exactly.

course. Yes, exactly.

>> All, right., And, I, think, well,, let's, share this again.

Finally, lesson four, the deployment.

For the deployment, we are deploying the fast API application using Rampot. We are adding the

using Rampot. We are adding the observability with OPIC. And finally

the Twilio integration. And let me show you the truly integration because that's really really easy the way that we have uh devised this. Basically in Twilio you

can see here that I have a trial account and with this trial account I have a phone number and the only thing that we need to do to test out our application.

If you want to build this for your business for example you will need to do this more properly buying your your own phone numbers on your own country. But

here under phone numbers you can go to 3ML apps and inside we have this app this fast RTC agent and this agent is

going to be connected with our fast API application., All right, and, in, fact, it's

application., All right, and, in, fact, it's going to be connected to this voice telephone incoming that you saw in the diagram at the beginning. Okay

[clears throat], it's, going, to, send, the 2ML instructions and uh it's going to spec back all the audio that is being generated by our TTS model. All right

and when you do a phone call, you will receive a very beautiful very beautiful voice with uh with the queries that you have.

Okay, by the way, you can see here that we are using Grock. This is something that you will do before deploying everything to Rampot. We are going to test out that everything runs locally

with coor etc. And then we are going to use the rampot enabled URL uh just to just, to, do, this., All right,, in, case, you

were uh wondering why there's enro in there.

Okay, and I think yeah this is uh the end. Uh to be honest I think we have

end. Uh to be honest I think we have been answering a lot of questions uh Jesus but I don't know if you have >> collected, any, any, interesting, question

during all the time that I've been doing this monologue but >> I, was, I, was, answering, people, I, don't know if, you, saw, my, messages, [laughter] >> leging some

this is a very good one a very good one why not VLM with uh LM cache uh so yeah very very good question and and to be honest we could perfectly have done that. Uh in fact one of the bonus

that. Uh in fact one of the bonus lessons that I was talking to uh to Jesus uh today was that after we teach you how to deploy with llama CPP and GDUF we were also thinking of doing the

same but with VLM also deploy it into into Ramport. Uh so yeah absolutely we

into Ramport. Uh so yeah absolutely we could we could go with with VLM. I also

love uh VLM for these kind of of purposes. It is true that for the GGUF

purposes. It is true that for the GGUF format I found uh better performance with lama CPP to be honest in terms of the results in terms of the of the quality of the

generated voices. Uh but if if we were

generated voices. Uh but if if we were about to use the the official the original model without quantization of Orthus tribillion uh probably I would go

with VLM and in fact I think it's a a pretty pretty good idea. Yeah.

Okay. John Michael, this is a good one.

So, are the weekly sessions going to be a hands-on walkthrough? Yes, they are.

Uh, from now on, we are going to get deep into the code. As I mentioned before, first of all, for the next lesson at least, we are going to cover everything you need to understand in

Jupyter notebooks. And sometimes we

Jupyter notebooks. And sometimes we probably will rely on some slides but very uh like very uh residually because we want to focus on the code and we want

to go line by line explaining how everything works and the different decisions that we have been taken throughout the the development of this course.

Okay., Uh, so, Jesus,, I, don't know, if, you have uh anything else you like to mention or some demo you want to to show or something.

>> I, think, maybe, I, can, just, show, you, a sneak peek of the kind of things that you can build with this project. And

yeah, should I should you just show it?

>> Of, course,, man., Of, course.

>> Let's, just, do, it., And, again,, this, is just demo time. So, you know, even meta at demo, [laughter] they they even fail. So, just beware.

This could be going a little bit wrong.

But I hope I tested like many many times today. So it should work amazingly. So

today. So it should work amazingly. So

let me share the tab. I think tab usually works better right Miguel >> to, share, the, audio.

>> Uh, yeah, to, share, the, audio., Yeah, I, was sharing the the tab. Yeah.

>> Okay.

>> Yeah.

>> Then, let, me, just, let, me, know, if, you, can see now the tab.

>> Oh, yeah., Yeah., Uh, stop., No, no, I, need, to I need to put it. I think we are >> Yeah.

>> Okay.

>> You, are, sharing, right, now.

>> All, right., Awesome., Awesome., So, this, is just a simple UI right because Miguel I think has made a really good point of explaining this thing fast RPC wraps the agent and actually answering you a

message you can take so we using langraph but you can use openAI directly you can use Google you can you can use any framework for building your agent but yeah so this is a different

interface right fast RTC wraps everything into this web RTC um protocol that allows us to do this voice realtime communication and this is just an UI that just takes the voice and just

streams the voice from the client right, from the browser. And I added something really cool that you're going to see. So, let us just have the

to see. So, let us just have the conversation.

>> Hello., How, can, I, help, you, today?

>> Hey,, what's, your, name?

>> I'm, Sophie., How, can, I, help, you, with, your property search today?

>> I'm, actually, looking, for, an, apartment with a terrace and two rooms. >> Error., Tool, call, validation., failed

>> Error., Tool, call, validation., failed

parameters for >> so, that, it, happened.

>> You're, welcome, to, help, you, find, that.

Can you tell me which >> validation, which, is, strange, but, okay then demo time failed and I tested it so many times this morning. Okay. Um

>> yeah, I, mean, that, that's, not, even, part, of the of the course, right? Uh it's like >> this, is, complete, like, a, exact, bonus lesson and then guys failed. Okay

>> let, me >> Yeah,, I'm, not, going, to, debug, it, now.

[laughter] >> Uh,, your, project, will, have, production ready code. So, I already uh mentioned

ready code. So, I already uh mentioned this, but yeah, we will uh give you all the blueprints that you need to put this

into into production. Uh, let me see how else.

>> Okay,, I, know, what, I, know, what, happened, but that will be in the next lesson.

So what is the primary Q&A forum uh substack? So yeah uh in fact let me show

substack? So yeah uh in fact let me show you the let me share the screen again uh here you can go to the chat uh to I mean

you can go to the neural maze if you put I mean I think Jesus already already shared this but this is the the neural maze like the

you can go here and and in here you can simply go to the chat. All right. If you

join the the new Amish chat, you can put any questions that you have in here. You

can put anything that you that you want in, here., All right., So, So, yeah,, go, ahead

in, here., All right., So, So, yeah,, go, ahead and and don't be shy. Uh so yeah, let's stop sharing and let's see any other questions in here.

The wiki sessions.

I have a question.

Okay, this is a very good one.

Okay, so I have a question. Is there any way we can solve network latency or do something about this because the LLM htts we use are all based on US server

and making voice AI agent using them from India adds up latency to be honest that that's a tricky tricky problem. I

would uh have you consider trying to host as many things as you are able to on your local servers like maybe hosting something close to your to your region

and trying to deploy your own models that that's basically what we are going to teach. So maybe it helps because it

to teach. So maybe it helps because it is true that if you are sending requests to servers on US for DLM for the ST

server and for the TTS server probably that's going to add a lot of latency and you you're right. So probably I would try to yeah I would try to replace maybe

the ST and the TTS models with uh with your own hosted solutions and maybe rely on the LLM for all the high intensive computations like tool usage etc because

probably that's going to be a bit hard if you want to do some call center or something like that. Uh but yeah yeah it's uh not an easy problem to solve.

Yeah.

Okay. I don't know who else is asking here.

All right. So, I think we are almost done. I don't know if you have any additional additional questions. So, here here John

additional questions. So, here here John Michael I did something with level apps.

Okay, this is a good one. How does this perform in low bandwidth scenarios? I

did something with level apps and that's the issue some of my users have. So to

be honest, we haven't tested with uh low bandwidth scenarios, but that's a pretty good uh a pretty good test. Yeah. Uh we

should we should test that out. Can you

elaborate on what happened with the with the user? I guess the the results were

the user? I guess the the results were not as expected and they were they were lacked with a lot of latency or what happened?

Okay, I think have a question.

Okay, I think we have answer.

Okay, look at this one. Uh, Jesus

you are muted. Uh, what should I learn to get my first job >> intern, as, a >> fresher, AI, engineer?, Yeah,, people, love

my my t-shirt things. Yeah. Uh, I mean my birthday was the 11th of November and they uh Yeah, they gave me this which is

a pretty cool t-shirt. Yeah, absolutely.

>> So,, yeah,, Jesus,, you, were, saying, what should I learn?

>> I, was, trying, to, I, think, I, fixed, the, a small tiny issue. So I can go again.

>> But, it's, better, to, to, make, the, surprise for future future >> lessons., I, think, it's

>> lessons., I, think, it's >> I, fixed, it., It's, just, like, um, a, small thing. A really small thing. So now it's

thing. A really small thing. So now it's working if you want, but um up to you.

>> Okay., What, do, you, think, about, uh, Rahul's uh, question?, What, should, [clears throat]

uh, question?, What, should, [clears throat] I learn to get >> what, should, I, learn, to, get, my, first, job intern as a fresher AI engineer?

>> Okay., really, really, interesting question. Um

question. Um um what should I learn to get my first job as a fresher engineer? I think

building things honestly. Um so just learn so what what to learn what to build. So I think in order to for you to

build. So I think in order to for you to show and to really be um someone that really cuts the attention of recruiters.

I think you got to show that you are capable of building things. So um

learning building and actually just doing the two things in parallel I think is the best way to to do it.

Absolutely. I mean I I agree 100%. You

know how the neural maze works. Uh we

love learning by building. So agree 100% with Jesus. Yeah. In this one just uh

with Jesus. Yeah. In this one just uh yeah build a powerful portfolio. I

always advise to build three endtoend big projects before 100 small PC's to be honest. Uh, and I could even, I don't

honest. Uh, and I could even, I don't know, take some of Jesus's projects or some of my projects and just try to adapt them to your own use cases something you really uh something that

resonates with you, something that you identify with and try to yeah, try to adapt that with uh to other use cases and put it on your portfolio. I mean I'm

surprised by the amount of people who have uh for example implemented AVA or implemented FO agent some of my of our previous uh projects

for some specific niche for some specific use cases and they they show them they share with the with the community and I think that's really positive when you when you want to get a

job. Yeah.

job. Yeah.

Okay. Uh, I just needed to to put this here. Uh, love your t-shirt, Miguel.

here. Uh, love your t-shirt, Miguel.

Thanks a lot, Himu. That's uh

appreciated.

Okay. Uh, so, hey guys, I think we are getting comments until like three three more minutes and then

we are we are leaving. But but yeah, so this is your last chance. Okay.

Okay. Can we can we implement this uh can we can we implement this any website? I I don't know exactly what you

website? I I don't know exactly what you mean but uh if you mean embedding this into any website then yeah absolutely that's, possible, at, the, end, of the, day

it's a radio UI the thing that I that we have shared with you. So yeah, you can >> but, and, even, not, the, great, UI, but, like actually the the actual WebRTC allows

you to build it and embed it anywhere.

So just with any front end you can capture the websocket and the WebRTC.

>> Okay,, let's, answer, this, one, because people seem to be interested. Uh what

are the skills required as fresher? So

for me it depends a lot um on what you want to what you want to become. You

want to become a like a new AI engineer you want to become an ML engineer. In

both cases I would say first of all programming language Python probably then I would go with some basic books

and books that give you the grounding basis of these fields. Tip Huyen has two two pretty cool books. Let me share with you.

So here is one. Let me share the screen.

So these are probably my my perfect my favorite courses. Okay.

For my favorite books if you want to to really understand what's the basis of of this field. So this would be for AI

this field. So this would be for AI engineering. this O'Reilly book and the

engineering. this O'Reilly book and the same for ML engineering. If I had to give you an advice, I would start with this one designing machine learning

systems and then after that probably go to the engineering and books are I mean for me at least they are really uh really interesting and they provide a lot of knowledge but you need to put all

that knowledge into practice. So I would try to probably go through each one of the chapters and try to find GitHubs or find projects that build something similar, and, just, try try, to, implement, it

yourself.

Okay, so let's put here uh the last question. Miguel, I follow you on

question. Miguel, I follow you on Substack and I want to know do you get ideas for the projects and what do you read or do which is byproduct which give

you ideas? So I guess uh how does my

you ideas? So I guess uh how does my like idea generation works to be honest there is not um

any workflow that I work with uh and I think Jesus is in the same point I mean someday I wake up and I say it will be amazing, I, don't know, to, create [clears throat], a, video, game, about philosophers and connect it with an

agentic back end and that's basically uh how my mind works and then once I have an idea that resonates with me. Once I

have, an, idea, like, this, one, for example like let's build phone calling agents I try to start designing in my head how it would look like what kind of use cases it will

solve uh what kind of features I want this system to provide and after that I just create the diagrams that you see uh typically in my stack okay I just try

okay I will need a fast API application I will need here a database I will need some way to provide the real-time communication that's how my my mind works. I don't know Jesus if you have

works. I don't know Jesus if you have anything to add in terms of your your mind.

>> No,, I, think, it's, it's, it's, like, that [laughter] honestly it's like that. Yeah. So, I try to build things that I that also resonate with me or that I will use myself and those are things that I that

I try to do it. Yeah. As well. But also

I think it helps a lot when people tell me, "Hey, Jesus, I I've seen this thing and wow, it would be really cool if you I, don't know, if, you, do, this, thing.", So

when I listen to other people, when I show things to other people, um they can give me some feedback or some other perspectives and that can help.

>> Yeah,, absolutely.

>> So, yeah,, that's, uh we need to stop guys, but uh come on let's let's get the the last questions in a roll. Uh so how do you get to know

these tools and connect with each other?

Most people can't connect tools like you both, do., So, if, you, want, you you, can

both, do., So, if, you, want, you you, can start Jesus and then I will give uh my answer if you want. [laughter]

>> Well, I, think, that's, by, digging, deep, into what you're doing. I think if you just I don't know if you're going to you like this thing and you like this thing and you want to put it together and you really don't understand how this thing works and how this thing works you're

not going to be able to connect them. So

if you go deep if you understand land graph well if you then understand fast artic well you're going to be able to say oh can this thing go together? Yeah

they can. And then you build it and you make it go together. Uh so yeah by just digging deep I think and not just cleaning in the surface.

>> I, agree, absolutely., Uh, I, mean, the, only reason we are able to connect all of these uh tools is basically because we have been fighting a lot of time with

these tools and we have been failing a lot of times connecting these tools and we decided to connect tools that were impossible to connect and we were like okay this was a mistake but you learn

how not to make these kind of mistakes in the future. So, so yeah, our ambition with the neuromaze with the the content that we built is to provide you the

guidelines and provide you the way this framework this mind framework so that you can avoid all the pain and all the repetition of errors that we have been

uh like making but yeah absolutely I mean if you start a new in something it's impossible that you know okay let's put here twin uh twillio and then disconnect with fast RTC and I mean

we've We've been trying a lot of different approaches to this project and this is the one that really works. And

the cool thing is that we have learned a lot of things to do when you are facing realtime communication projects and also a lot of things that will get your application directly to the trash. So

so yeah, that's the the cool thing about testing and and building like like crazy. So

crazy. So okay. So I'm just going to read the

okay. So I'm just going to read the three last articles and we need to go.

So let's say you are recruiting an intern for your project. What things

would you like to have in Kim? So I can start here. For me it's not as important

start here. For me it's not as important as you know the different techniques for rack optimization for chunking. Uh you

know everything about recent agents, you know all the LLMs that there are. I

don't care about that. The only thing that I want for you is that you're willing to learn and that you're willing to work hard. That's basically what I want. Of course, I mean, I would want

want. Of course, I mean, I would want for you to know Python, to know pretty basics about AI, about ML, deep learning a little bit. But for me, it's much more

important that if you don't understand something, if you don't know how to build something, you research, you put a lot of effort, you are independent.

That's the kind of people that I would like to have in my team. Yeah.

Completely that was a really good answer. I think

for me it's the same and with AI and your speed advance of AI it's just like I think the hard skills or just this hard knowledge about I know this small thing about this paper that I read in I

don't know in 2019 but I mean that's all right but you're going to face bugs you're going to face just error connections you're going to you you need to be you you need to have that

ownership that independence to really um just work on things by yourself. Of

course, in the beginning that's more difficult, but slowly and slowly you will you will get that for sure. And you

will get there by just not surrendering [laughter] by actually doing the work.

Doing the work and then you'll see that everything just gets solved easier and easier.

>> Absolutely.

>> And, there, is, this, word, in, in, Spanish.

How do you say that?

>> Critical, thinking,, I, guess.

>> Thinking, I, don't, what, I, mean, is, just that Yeah. That you got a you got an

that Yeah. That you got a you got an Yeah. It's difficult to explain because

Yeah. It's difficult to explain because I'm critical.

>> Yeah.

Okay. So, uh two last ones. So, we can change the text tax, right? Absolutely.

>> Critical, thinking.

>> Critical, thinking., So,, yeah,, you, can change the the text tag. Uh no problem.

And the last one that would be nice if future demo using a phone in addition to on web. Absolutely. Uh I mean, we just

on web. Absolutely. Uh I mean, we just we just didn't want to show you everything in lesson zero, but absolutely. I mean this course ends up

absolutely. I mean this course ends up with an agent connected to the phone. So

probably on lesson one we will start just connecting everything to Twilio but just, with, your, local, setup., All right.

Probably this lesson is going to end up with a connection with um with Twilio but using Koko and Moonshine. Okay. So

so yeah absolutely absolutely.

Yeah. related to critical thinking here.

Strong op says uh lateral thinking is good too which I I [laughter] agree that's a a good point. A good point sure >> yeah, yeah, pretty, good, point >> but, yeah, but, the, answer, is, just, not, so

much the hard skills anymore.

>> N, no no no, okay, so, it's, been, a, real pleasure. Sorry if uh we had some

pleasure. Sorry if uh we had some problems with Substack live at the beginning. It's our first time doing

beginning. It's our first time doing these live sessions. Uh please provide any feedback. Uh I mean try to create

any feedback. Uh I mean try to create constructive feedback if if you can.

>> Yeah., Be, nice, with, us, and, provide, any feedback any things that you would like to see in the future. Things that we have done right, things that we have done wrong. Uh because we will try to

done wrong. Uh because we will try to improve for next Sunday. So yeah

>> it's, been, a, real, pleasure., Uh, we, really enjoy uh spending this this time with you. It's been really cool and I hope

you. It's been really cool and I hope you have enjoyed this lesson listen lesson zero and mostly I enjoyed that I will see all all of your faces and icons

in next lessons. So yeah

I don't know if you want to say uh goodbye as well Jesus to the community.

>> Yeah,, I, really, enjoyed, this, session, a lot and I'm really looking forward to see you guys in the next one. I I really like I think I enjoyed a lot this part of like talking with the people and just getting the super interesting questions

from people and just answering. So this

was also really cool. [laughter]

>> Absolutely.

>> Yeah., Thank, you, guys.

>> Okay., Thank, you, guys., And, uh, before, we leave, remember that you will receive all the free subscribers and premium subscribers of the neural substack will receive an email containing all of this

recording in case you miss it anything and also the slides so you can download the slides if you want. All right. So

have a nice Sunday, rest of your Sunday