"We Killed RAG, MCP, and Agentic Loops"

Okay, cool. Hello, everyone. Welcome back to TwoSet AI. Today, I'm sitting down with Paul Eustine, author of the best-selling LOM engineers handbook. If you have it with you, show us your book. Yeah, yeah, awesome. This book and he's also the founding AI engineer at a San Francisco startup. ⁓ He also has a very famous

Yeah, yeah, I have it here. Of course.

newsletter called Decoding AI, where 40,000, more than 40,000 engineers are learning how to actually ship AI that works. Paul, thanks for joining.

Yeah, thanks you. Thank you for inviting me Angelina. It's a pleasure to be here.

Yeah, awesome. I mean, you've shared so many interesting things about yourself and I read your blog post. I have to tell you that you gave me a few laughs reading your blog post. You even write like, hi in your blog post. Do remember? The one you wrote about, we killed rag MCP and that one, I think it's very timely and very useful.

In which one more exactly? ⁓ yeah, yeah, that one. Yeah, it's a controversial one, but it's true. I don't remember the haha, but I like to this more informal type of thing.

Yeah. It makes sense. Yeah, it makes sense. You did you put a high in it. All right. Okay. I you know, you told me last time when we met that you haven't done a traditional job interview in five years. But you're now a founding engineer at a San Francisco startup. Like how did that happen? How did you get a job?

⁓ I don't think that it's really five years, but let's say three years each, maybe more ⁓ But I guess that's not the point ⁓ Yeah, so basically I realized that

Three years.

like this lead code job interview path is kind of hijacked because there's a lot of competition, you know, if you go in that path, are many, many smart people who basically will compete with you on the same job positions. And at the same time, even now with the era of AI, like it's just a numbers game. So I realized this like I realized this like five years ago, and especially because like I'm coming from a different background, like from Romania, where I don't necessarily want to apply to positions located here in my country, because there are not many positions and the ones that are here are not really that interesting. ⁓ I realized that I need to go more into the like contracted freelancing route and Very quickly I realized that while doing these interviews, it was really impossible to say what I know in like 30 minutes or a couple of minutes or whatever, right? It's just the game is not under your advantage. And also in the AI space there are so many dimensions ⁓ that people can actually ask you questions that is impossible to be prepared and not waste time.

Mm-hmm.

getting prepared for these interviews and getting asked questions that don't really matter. They're just like robotic questions and robotic definitions that don't really matter. And I'm sure that many people ⁓ resonate with this.

you Yeah. Yeah. The sentiment is definitely shared. It's been a little too long. Three months to six months to prepare for lead code. And then, you know, there are like robots interviewing you as well. Like AI is... You didn't know? I heard from my friend.

Yeah. Yeah. Yeah, to be honest, I don't know. I really don't know anything about the interviewing process anymore, to be honest. But that's why I started outputting everything that I know on the internet. That's actually my first drive on how I started creating content online. I did not have any dream like being a content creator or things like that. I just love to learn.

Mm.

and I did not had really a great time. saying what I know in those 15, 30 minutes. And I just started outputting everything that I learned online. that was like basically building in public is magic. I would say, I mean, there's no secret recipe. You just need to be excited about what you learned and just share it online on LinkedIn, on your blog post, on YouTube or whatever makes more sense to you. And if you keep doing that, then you're actually I mean, something relatively interesting to say, like you don't have to invent the wheel or be a genius or anything. just have, the thing is that many people are going through the same process as you are going while learning. So they will be very interested in what you have to say, even if it's not new, just the experience itself is valuable.

Is that how you lend your job?

Yeah, mostly. And the thing is that because my jobs are framed as contracting positions or freelancing positions, I'm not going to the big tech route. Most people just come to me and these interviews are very basic and they don't necessarily ask me like... ⁓ definitions like the classic interview to ask me questions to see if I know things because it's obvious that I know these things because of like my online portfolio. And basically they just asked me how can I help is more a business oriented ⁓ type of questions like how can I help in how much time can you deliver in like what are the costs for you to deliver these things? So it's more like framed as a service.

Right. Yeah.

then as a job and if you start selling things like this, you basically, you don't go the interview route. It's just, you're just seen differently in the team setup, let's say, because the reality is that if you go this route, you're not necessarily part of the team very often. It depends a lot from company to company.

Yeah. Right, yeah. I think it's an underrated path, like career path for a lot of engineers, right? Because ⁓ the traditional interview path is more like filtering. They're filtering from a large number of talent pools. And what you're getting by publishing and showcasing what you're building in public anyway, People come to you to ask you to help solve their problems. Right? Can you help? It's more real. It's a real thing that they need. It's not like interview. I'm interviewing these people or I need some help. Right? The framing is totally different.

Yeah, yeah, exactly. Yeah. Yeah, exactly. That's the trick. And I guess it's like somewhere in between like being a like the classic interview where you're basically just want to be an employee and more the entrepreneurship. It's not necessarily 100 % entrepreneurship because ultimately you're just having one client or two clients, which ultimately it feels like a normal job, like internally.

Right.

It's kind of the same, but how you sell yourself and how you position yourself in the market is completely different. And again, I don't think this works everywhere. if you... What you strive is big tech, most probably it won't work or it will work in very siloed situations. ⁓ If they really need... They don't want necessarily to fully hire an expert on something that just needed for a specific project that can exist, but it's not like the, I don't know, the standard big tech dream that many, people have. And personally, I don't like corporate. I don't like that fact. So it's just a personal fit.

No, yeah. Yeah, that makes sense. I feel the same way. after I worked for bigger companies for many years and then I came to realize that what I want in my career. yeah, it's an afterthought. Like I didn't plan it. When did you when did you start writing? You publish a lot of content.

Actually, I started quite a while ago, I think ⁓ four years or so, something like that. But in the beginning, it was very basic. I actually started writing a paper out of my master's ⁓ degree, my project. And I kind of got pretty bumped by the whole academic process. Again, I did not really liked it and I found like this more social media medium, sub stack kind of writing a lot more fun and entertaining. And I took that path around four years or so. But at the beginning it was very basic, nothing serious, but as things got traction more and more, I started to do it more often. And now I write like on average one article per week or something like that.

and we hope. You also build a large following on LinkedIn. mean, when you first started posting on LinkedIn, let's what did you think was going to happen? What were you hoping for? yeah.

I don't know. Like I had zero strategy. I just said, okay, I like to learn. I like to learn new things. I'm really passionate about software and AI. And I just started posting randomly without zero strategy, zero. As I said, I just wanted like in my mind, I was, I was imagining that I would create like exposure and I will get easier.

Thank you.

like these contracting jobs. Like I would have never imagined that we'll end up like in a content creation education type of thing, which is now like, I did not plan for this at all. It just intuitively ⁓ got here.

Yeah, yeah. Was there a moment that you realized that the social posting was actually working? Was there a moment that you feel like, there's a pool?

Yeah, actually, I think the biggest moment was when I managed to meet people from all over the world and this slowly opened a lot of doors. I started to meet people from all over Europe, then started to meet people from the US and got into this, slowly got into this AI community, which felt amazing. I had these amazing conversations. ⁓ I felt I was there in the game. was a lot easier for me to ⁓ basically stay up with everything that's happening inside the industry. yeah, to be honest, it's also very engaging to create these LinkedIn posts and see how they perform, how they go, what works.

I really like your post because you're very good with visuals. I really like your visuals. feel self-explanatory oftentimes.

I'm a visual learner. For me, it's very hard to my thoughts into words. But visually, it's a lot easier. It's very intuitive.

Yeah. Is there a tool that you're using for creating those diagrams?

Yeah, I'm using Canva ⁓ for now and I'm still doing them manually like a caveman for now. I kept trying to automate this with all these AI tools, but I don't know. It feels like they take a lot of the magic out of it.

Okay. You're doing it. Okay. Yeah, yeah, yeah, I've tried different things, but I still like yours best. I thought, this looks really nice. I get it.

Yeah, I put a lot of effort into that.

How long does it take for you to create one on Canva?

Well, it depends on the complexity, on the more complex ones sometimes I say like 30-45 minutes.

Yeah.

It depends, it can take me from 5 minutes to 45 minutes, it depends a lot on the complexity of it, on my inspiration, on everything.

Cool. Let's talk about what you're building at the startup. Are you doing full time right now?

⁓ Actually, I know I'm part-time otherwise. I did not have time for everything I want. ⁓ So we were building a vertical AI agents for finance. Like we kept pivoting many, many times because of how the market performs. But yeah, basically we are working in this

Okay, okay. What are you

financial advisory space where advisors need to stay in touch with clients. And there are many things that, many steps that are very like document-based, report-based, document-based, that's a lot of manual work, which can easily be automated basically with... ⁓

Mm-hmm.

It's not that sexy, know, but ⁓ it's exciting. Like the tag behind it, it's exciting because ultimately you need to a system similar to any other domain. And I think this vertical AI agent space, it's actually pretty... Like how you build them is pretty much the same with small variations from application to application. In my mind, they're like the new web apps. websites of. Probably I expect that these processes will be a lot more templatized as the years go.

Hmm. Yeah, yeah, that's very actually that's enlightening. I mean, it sounds like this is just a financial advisors are providing this customer service to their clients and then you're just automating that serves. Right.

Yeah, in this particular use case, yeah. But I was saying that I expect that all these vertical AI agents applications, regardless if they're like in finance, medicine, e-commerce or whatever, I expect that many of these applications will be the same as time goes. And how you should build them will become more and more templatized and we will find a blueprint that works like similar to how at some point people realize that know these MVC crude systems are the best to build like web normal web servers.

to the next. Yeah. Are you seeing these templates are going to be productized? So that non-engineers or non-developers can, in the future, you know.

Not necessarily templatized, but like... ⁓ frameworks will find more like go-to methods on how to build them. Probably with all these coding tools that we have today it's hard to tell, you know, how... ⁓ how easy will be for non-technical people to implement them? That's a very ⁓ hard question to ask, to answer right now, in my opinion. It's hard to guess, but what I'm saying that just like the framework, the mind map that you need, like the steps that you need to take to implement this type of things will be more and more like known, right? Because...

Yeah.

I know one year ago when we started building this, we had no idea how to build vertical agents to like properly do rag on multiple data sources like documents, images, text, or how to properly do like a gigantic rag on top of them, how to plug in at some point audio into all of this and how to expose this to multiple integrations and all of this and... That's basically my article, what it's all about. We tried all these fancy algorithms, hoping it will work best, but we realized that... Usually the answer was more in simplicity than following all these fancy algorithms. But it's hard to find this simplicity, know? It's the hardest.

Yeah, I read your article. ⁓ I mean, as engineers, ⁓ you and your team members are, you still struggle with, like when you start building this, you still struggle between the decision of, do we use the fancy latest thing or what should we use? That's an interesting observation that when I was reading your article, I felt like, you guys struggle too. And you guys, right?

Yeah, I think everyone, everyone struggles.

What went wrong with rag in that application? You said you killed the rag.

Yeah, we killed it for not entirely, but for many scenarios, we ended up not using it because for example, in our particular use case, like every user could like through the UI and UX that we've built could easily like solve, silo the information it has in one chat at the time, right? So that's like how we guided the user. So basically, let's say that the memory itself that's like loaded or you can do rack on top of it siloed by default. And after we really understood our business use case, we realized that For this particular use case, like for most of our use case, if we load all the data that an advisor has into the context window, it will just sum up to 64,000 tokens at maximum if they load everything they have. And in these scenarios, it's just easier just to load everything into the context window and just pass it to the model. And basically you don't need to do drag to query only what you need and all of that. And why is that actually ended up doing rag actually ended up costing more or being slower? Because when you do rag, you have this retrieval problem, which is actually pretty complicated that you need to pull in the right data. And the model, if it doesn't find the answer in the data it pulled, it will start to query the memory again and again and again until it finds. And this is if you're lucky enough and the agent realizes that he doesn't have, has the right data and it doesn't start to hallucinate. So it's basically this agentic loop, which is like a zigzag, which just. adds extra latency and potentially even extra costs because you do a lot many more LM calls to answer your question. And if you do like, if you just load everything inside your context window and you answered your record based on that, yeah, it takes longer and it costs more per that LM call than doing rag. But ultimately if you sum up everything, it's faster and cheaper.

Mm.

Many times, not every time, cheaper.

Hmm. That's an interesting ⁓ insight. You're saying that if I add it, so, so the, so the use case is fairly simple. You have some clients data, like some documents and some data, right? So you need to pull out to answer some questions and you're saying the amount of data is not that much pertaining to one customer anyway. So we can actually have everything in the context window versus when you're using rack to do retrieval.

Meh. Yeah, exactly.

and you have agents deciding, making a decision, whether, okay, did we get information or not? Then, then this is like talking to a human, right? If I, if I asked my intern, like, you know, get this file and then make sure, make sure double check if it's actually the stuff that I want, the data I want, most likely, I think the human will go back and double check, right? So, right, right. ⁓

Yeah, that's a great analogy.

Yeah, because they want to be helpful. when they double check and they go back and then do it again, which that recursive loop could happen, which increase your cost and latency. That's what you're saying.

Yeah, yeah, exactly. And that's what I want to highlight. Technically, the problem itself is not fancy, like it's a simple problem. Everyone knows that, but when you frame it in a concrete business problem and when you're put to actually solve a real world problem where you can have so many decisions, doing the right one, it is hard. Even it's like choosing the simplest solution.

Right.

It's sometimes hard to see through the fog, you know?

I want to ask you like, why didn't you guys choose the simple solution? I feel like this is a simple solution. How about we just give everything to the LOM and then let it decide?

Well, we actually because we pivoted many many times. ⁓

Okay.

We started with a more complex idea. Like we did not have clarity on our business problem since day one. And I think that's the biggest learning like to actually frame really well your business problem. But sometimes the issue is that you don't know yourself, your business problem, like this particular context. Like we did not know in advance that in so many scenarios our context window will be. our data will be small. But from what I've seen more and more like with experience is that actually these vertical AI agents, in many scenarios, you actually have small to medium data and not big data. And all these fancy rack techniques often are framed for medium to big data. And I also built a

will be enough.

this agentic engineering course which we recently released where for the capstone project we actually did a deep research agent and like an agent that writes professional articles. I applied for example did this same pattern here after I realized it and the thing is that like to write a professional article you often need as references I don't know five, six other articles. And your first thought is that, yeah, if I want this agent that writes this article to have access to all my knowledge, need drag, like it's obvious, but no, it was just a lot easier to put everything in there. And that's it. With some smart prompt engineering, maybe some context engineering, like how to put everything into the prompt, how to filter your data properly, not...

Mm-hmm.

put everything inside your problem. It's easier to just take everything and do some filtering and do this retrieval. And I believe that many, problems can be solved this way. Or just do some simple SQL queries with some metadata filters and sorting and things like this. And it's just easier, you know.

That's a really good takeaway. mean, I get it now. So I mean, the team started with this assumption, right? There may be larger amounts of data that you need to work with for this use case, but you didn't know at the time, basically.

Yeah, yeah, exactly. And I think that many teams, many people start with this assumption or don't even realize that they start with this assumption.

Right. Right, right. So I guess the Silicon Valley, you know, talk to your customers. Still works. Right, work with your customers and then you'll know. Right?

Yeah, it's great. But the thing is that we did that, but especially in a startup where you pivot so many times.

Right.

Your initial customer can be different from your current customer, you know. Everything is so noisy and that's why it's hard to take the right decision every time.

It's also true. Yeah, yeah, I see. Yeah, I think that's the true and the biggest challenge in light building something, right? That's why I'd make it hard.

Yeah, like to take the direction and know what to build is the hardest, especially now when AI can like help you code and build stuff. This is, think that the most important skill you can have.

But. Right. Yeah, yeah. Do you think Rag is dead?

No, ⁓ no, but I think that is is like an algorithm that you should use when it makes sense to use just that.

When does it make sense?

Well, basically, based on what we talked so far, when you have like medium to big data, that's it. Or when you want to be your super constrained on your context window, right? Maybe you want to use a smaller model, a cheaper model for whatever reason, and then you need to be very aggressive on what you put in there. And then Reg is still a good solution. basically the ratio between your data

Mm-hmm.

and your context window. ⁓ I think that's a good formula that you can find. Do I need rag or I don't need rag?

You know, that's a good one. think that could be my LinkedIn post today. Right. So creative. Yeah. And you also like in your, I mean, in your experience in this, this use case, right. You also said you, least from the blog post, you said you regretted using MCP as well. Well, everybody is so excited about using MCP last year. Right. You should always use it. What happened?

Yeah, I just come up with it. Yeah, I also think it's a good solution. Like I still use MCP, but again, everyone was using MCP everywhere. And I think that's the wrong. I think that that's actually the take of the article. That if one solution is hard, you shouldn't put it everywhere. Like if you have a hammer, you see nails everywhere.

Yeah, yeah, yeah.

This is again what happens now with AI agents. How should I solve it? AI agents, doesn't matter what you want to solve, AI agents, it's exactly the same.

yeah. Yeah. Right. Yeah. Yeah. Yeah. And like, if I, let's say, if I want to build some, some agent tech system or whatever, you know, AI application, and then I come, come up to you and ask you, Hey, Paul, here's the thing I'm building. What's your bet on the chance that I need, I actually do need MCP in my stack.

I think if it's a tool that's very dev-oriented. dev-oriented or like for heavy users that... So basically I see MCPs plugins, let's say very intuitively. And you can either use these plugins in dev tools like Cursor, Cloud Code, or people that use Cloud or similar tools in other setups, because now people started to use Cloud and tools like this in creative work, productivity work or whatever. Basically MCP is a plugin, but we were actually building a product and we needed integrations with other providers like email and I cannot go into too much details, but we needed integration with many, many, many providers. And we thought that every provider provides an MCP server and why shouldn't we connect to them through these MCP servers? The thing is that our MCP server work was sitting on our server and it wasn't hooked like directly to the client, like to the user. We, it were connected on our server and ultimately we just ended up like writing custom MCP servers for each of our clients for all this integration. And we haven't used any of them because there were two rigid or custom use case, they just did not have, weren't updated enough, or they just did not add yet an MCP server. So for example, now I still use MCP servers because I'm a heavy user of cursor and cloud code. And when I want to write myself ⁓ a custom code to hook them to these orchestrators, ⁓ I still do that to MCP servers. And now with the skills, ⁓ cloud skills, That's a different path. So you have two ways to do that. But ⁓ on our end, we had to build everything. The orchestrator, how it uses tools, how it integrates to other tools. So just by passing all these extra protocols and writing everything on our end, it was easier.

Mm.

I hope that that makes sense.

I mean, you have to write the functions itself anyway, right? Because the MCP.

Yeah, yeah, but the for example, let's say that you want to connect to Gmail is a common example.

to know.

They provide an MCP server, right? I think so, at least. And you can connect and access like through an agentic orchestrator Gmail, through an MCP server that's run on their end, or you can just like implement their API, which is like a simple crude API and put tools. ⁓ Basically, you wrap this API through tools.

they did, okay.

Which is ultimately the same thing, but like if you actually write the code on your hand, you have full control and full customer customization on how you do it. And you can do it exactly how you want. You're not like, I don't know, lost in abstractions and things like that.

Right. What do you think of the, so what do you think, what are your thoughts about this MCP landscape then? Do you think it is going to be useful? Or like based on what I'm hearing, you think it's not customized enough, basically?

Well, I think it is like, for example, if I use cloud code and I just want to hook to my Gmail to do some basic stuff, like read my emails and things like that, I still think it's amazing. Like I don't want to implement some custom code when I want to hook to every integration with cloud code or cursor. For this use case, like when... it's added as a plugin to other existing tools that I don't have control over as a developer, then it's amazing. But when you build an application where you have full control over, then I don't know. I wouldn't say so. I would just skip it, but who knows?

So it's use case, basically. It's use case dependent, how complex your problem, how customized you need it to be.

Yeah, yeah, but that's again, ⁓ that's what we thought. But because the thing is that when you build like your custom orchestrator, if you want like to either choose, if you want to build your MCP or not, then you have to integrate MCP protocol in your application, which adds those extra layers of abstraction, which make the code more complicated to understand.

Right.

and add more extra dependencies and things like that, which you don't want, especially in this world of AI, want to keep it as dependency free and simple as possible so that AI can actually write good code for you and not get lost into so many abstraction layers. If you start using MCP just with the idea that maybe I will use it, maybe I will not, you will just add this extra layer that adds useless complexity. then it's not like a gray thing, gray is more like yes or no. Do I use MCP or not? And so what I see, I think this will help more. MCP is more like a way to serve your code. I think. And it's like, instead of serving as a fast API server, RESTful API server in general, ⁓ you serve it as a MCP server. And again, if you're building the server itself, of course you can connect to other servers, but if you have full control over the code, I suggest, and I recommend not to ⁓ use MCP.

What do you think of the very hot like vibe coding tools out there? I mean, including, you know, lovable and, you know, those kinds of tools that kind of enables people who are non-technical. ⁓ I got to believe a non-technical person wouldn't understand your blog post that specific one, for instance, right? They might not go that far understand how scalability ⁓

Yeah.

look like and what's needed. Yeah, but how do you, what do you think of today's like hype on these vibe coding tools?

⁓ Yeah, personally, on this I'm pretty bullish. ⁓ Like, again, it's hard to say how far they will go, but I will give you an example on my experience. Like, I don't know how to code front-end necessarily. Like how to do designs and all this. And for example, even for me, like a engineer, they can just extend my skills because I know enough front-end and design skills to vibe code this app, if that makes sense. And like this, I'm really confident that I can build end-to-end applications with these tools. Otherwise, it will take me so much time to do them manually or even surgically using LLMs to build specific parts. or off the front end. I started to be more and more bullish on Vibe coding and I was thinking to myself like less week or so that many people complain about Vibe coding that it creates like messy code or code that doesn't respect like clean code or good architectural designs and things like that. But the thing is that these designs were made for people

Mm-hmm.

to avoid human error. And maybe the AI will not need them anymore. Maybe that's how AI actually writes code and we should live it. We should not fight it. should live it and do its own thing and just adapt and actually create some layers, other ways on how to evaluate and make sure that the systems work properly. Like we should... In my mind, we shouldn't necessarily force AI to think like us. Because that's pretty rigid, like why we should do that. We should just like find a more intuitive and natural way in ensuring that what the AI builds, what we really want to build then. But I'm not saying that like engineering is dead. I'm saying that it is just like the game is changed. you should be careful at different things. Like you should probably still decide how your system looks like, how like, I don't know, your data models look like, how your data flows, how everything scales, the input output, the business problem that you solve. Like you're still the one in charge that solves the problem, but you're not necessarily the one in charge on how the code looks behind the scenes. Ultimately, now we write Python that's compiled

Mm-hmm. Mm.

many, many times until it reaches bytecode. Do we read the bytecode to make sure it's clean bytecode? No, we just trust it. And we can see AI generated code like just a compilation step. don't know. I'm just saying.

That's a good point. No, no. Yeah, this is very interesting. You are saying that it's like, you know, we have to accept the younger generation speaking a different language and then we just work with it.

Yeah, yeah, that's also a great way to put it.

is And it becomes lower level and you don't care that lower level anymore. And maybe it's another level that we have to reach and then we haven't developed it before. Right. Maybe it's ⁓ product level. I don't know.

Yeah, exactly. Exactly. And we need like a way to ensure at the next level that everything works fine. And for example, one level is like AIEvalues, which basically ensures from a data point of view that your product works fine. Like you don't even need to write like the classic unit and integration tests. You just like from...

Mm.

use a data approach based on specific inputs and outputs and ensure that your system works as expected. And me personally, I think AI e-velopes will be a a huge domain in the future because if AI can code, then the only thing you can do is to evaluate it. Yeah, exactly.

Yeah. Yeah. Yeah, make sure it works. Right. Yeah.

and to create a very good signal back to the AI on what to fix and how to fix. And also it's extremely important to know how to prioritize what to fix, because we shouldn't expect that we have infinite AI. We know now that with I know cloud Opus 3.6 is amazing, but you run it a few times and you run out of tokens. So it's also limited resource. So you need to know how to prioritize, what to be fixed and how to be fixed and things like this. So we kind of play a final game which we should take as such.

Good point. The evals are the new muscles we need to build, which is a lot more stringent requirements compared with the old traditional ways of the unit testing, regression testing. We need to do a lot more, right? Because it's not so deterministic anymore.

from Yeah, exactly. From many point of views they're pretty similar, like with unit tests, regression tests, there are big overlaps, but as I said these elements are not deterministic anymore. You also work with a lot of unstructured data like text, images, video, and you cannot just do an equal operator between two pieces.

Mm.

of text because they are not identical but they are the same if that makes sense, like you never expect. I mean, we know it's that you can reframe a sentence in many, ways and it has the same meaning. yeah.

Yeah. ⁓ Tell me about your new course, right? So you touch, we touched upon a little bit about it. When is it releasing? Is it on May the?

We actually just released it ⁓ yesterday. Just in time. Yeah. Pretty excited about it. Can you repeat the question please?

Yeah, just in time. How long is your course? How long is it? How long is your course? Is it a one day?

At the moment we have 34 lessons and it's placed on three modules. The first one is on the foundations of AI agents where we build from scratch all kinds of workflow patterns, agentic patterns, multi-model patterns and things like that. And the next two modules are the module 2 is where we build two ⁓ capstone projects.

Okay.

which then we connect together, like there are two agents that work independently and also work together or with a human in the loop. And the third part is where we build AI evals, monitoring, we deploy them to production and things like this. And we will keep adding lessons as time goes, we will update it as many as we can.

Is this live? Is this live cohort?

No, no, it's self-paced with articles, videos, notebooks and the

I'll self paste, okay. Yeah, did you post? Did you? I didn't see your post. You must have shared it on LinkedIn, did you?

No, no, we are actually doing a more long-term ⁓ marketing strategy, but I can show you the link right now.

Okay. Yeah, yeah, yeah. Let me know if you have any timeline you want me to work with that can help you because I think that when the video rolls out, so this is on the side that the editing might take three to four weeks now. But if you have a timeline, I can try to release it sooner to help you market your core.

No, I think three weeks is perfect

Okay. Okay. Yeah.

tricks is amazing. Can I share it in the public chat here?

Yeah, yeah, go ahead. Go ahead and I'll add the link in the description. Do our audience get a discount from your class, from your course? Give me a discount code if you can.

⁓ We are pretty careful with discount codes to be honest We have we did not gave ⁓ anyone

Okay, it's okay. No worries, no worries. Yeah. okay. No worries. No worries. If available, I'll definitely share. If not, no worries. It's totally understandable. Yeah. So ⁓ I want to ask a little bit more about your course. Can you share what are you learning in your working experience and your building experience for your clients that will actually go into this course?

Yeah, so basically the goal of this course is not to get super theoretical, but to build an intuition, like a strong intuition on how agents and LLM workflows work. Because in reality, when people say you're building agents, you're either building workflows, agents, or somewhere in between. So we build a strong intuition on what is what, what patterns go ⁓ where, ⁓ and how actually to combine them. And then we build these two capstone projects mostly from scratch. We don't use like those ⁓ agents from LandGraph or things like this, because we want like to actually understand how things work behind the scenes. And I also am a strong believer that you shouldn't use agents from frameworks in your code, even in production code. they are great for like

Mm.

quickly prototyping and seeing if something works. Yeah, they're amazing. But if you really want to build something like a custom project that's like production ready, you'll most probably need to build a lot of custom code, custom orchestration logic, which those frameworks don't provide. So we actually need this for like to actually build a really well applications. But at the same time, I also think that frameworks like LandGraph are amazing, like to give you a ecosystem of utilities and memory layers and connectors and like retries mechanism and everything around that. ⁓ So in this framework still have a lot of utility in that space. And for example, that that's why we use the Niagara functional API, which lets you like build whatever you want in plain Python code, but they also come with this memory layer checkpoints. So you can very easy debug, resume your code and things like that.

Thank

So we, we exposed it as an MCP server, ⁓ because we want to like to leverage existing UIs from, from like cursor or cloud code. So instead of like building our own UI and own application, we, took it like a productivity dev tool. So it's plugged to these tools. And we, I also liked how we designed this, ⁓ human in the loop system.

Mm-hmm.

So for example, for the writing agent, we took a similar approach to coding where basically after we have the first article, we implemented editing loops where you can select what text you want to edit and you send it to the agent and it responds with a diff where you can very nicely accept or reject what you like or not exactly like coding and things like this. And also I'm really excited about the EIEval section where we implemented very specific custom business metrics because in reality ⁓ it's not good practice to use this generic metric like hallucinations, helpfulness and things like this because they are not helpful. They don't show you any business value. You have a score of hallucination of 3.8. What does this tell you? Not much. You can just assume what it tells you, but can you go exactly to the code and pinpoint to what you should fix? No. So you should, again, be very careful on how you build this specifically for your data in business use case. And we'll show how to monitor all of this, track it, plug it as offline levels. build some LMGRGs that are actually calibrated to the domain expert and also use them as online evals and lot of things, many things around this base.

So you actually have a of like a start from scratch course, not using frameworks, but ⁓ help people to start building AI agent systems and together with monitoring from beginning to end. ⁓

Yeah, exactly. Yeah, exactly. ⁓ I have a saying that I keep using is I like to teach people how to fish and then they can do whatever they want. Like if they understand the decision process, because we kept talking about this, if they understand the decision process and understand what to be used where, then the tools are not that important. You can either write from scratch, pick a tool, pick an algorithm. That's actually easy now.

haha

With AI, it's easy to implement this to some extent, of course, but knowing how to pick what, where is the hardest part. And that's what the core of this course focuses in. And you can be like a complete beginner when you start, and we take you from ⁓ beginner to advanced.

Yeah, yeah, I think you'll insert. ⁓ got it. That's really nice. So non-developers can also attend.

⁓ No, you need like some Python and LLM knowledge before. It's not like a complete beginner. You need to be like a developer, engineer, data engineer, data scientist or some knowledge before. Like when we go straight into it, don't lose time with the basics.

Okay. Okay, okay. Some knowledge, yeah, okay. Got it. Okay. ⁓ You know, if a student comes up to you and then because there are, you know, many AI agents courses out there, right? How if a student comes up to you and asks you like, what can I do after I take the class? ⁓ What's the before and after that you can tell them?

Well, you will have three projects that you can show, like two that are built with us, where they are basically what I explained, I don't want to repeat, but... also you need to build one project of your own. So basically you've built a full portfolio that you will put on GitHub and you can show... to your future employers or your future ⁓ or if you're going the contracting route or online route, your future audience, who knows? So, and they're not like basic with you, for example, notebooks just for teaching. These are full fledged Python projects that are like shipped with UV, actually shipped to GCP.

Mm-hmm. Mm-hmm.

⁓ with authentication user layers and everything. So basically you can just take this framework and go at your job and implement agents for your company.

Do you feel like in today's job market, this is like, you're building your own portfolio of these AI projects is necessary for looking for a job or looking for freelancers?

⁓ I think they're the default of what you need to do to actually to consider applying somewhere, right? Like if you don't have this

What about university? What about like graduating from a university with a master's or bachelor's in computer?

I know you eat. Yeah, maybe if you graduate from like Stanford, MIT and this top universities that counts. But if you like graduate from a normal university, I don't know. It's again a default. Like you need to check. Okay. Degree, project, and then we talk. That's how I see it.

Yeah. Right. Right. Okay. Yeah, that's a good takeaway. you know, the engineers is like designers, you kind of have to have a portfolio, like a pamphlet, a book of portfolio to show. It's just like,

Yeah, exactly. exactly. Otherwise, especially in today's market, if you don't have like a minimum of proof of work, then it's hard.

Right. Yeah, yeah, it's hard. And when you show a lot of your proof of work, like how you did it, then you get people reaching out to you, right?

Yeah, exactly, but you need to put in a lot of work for that.

How much work is that? How many years of work is that?

Well, ⁓ I started coding pretty heavily like 10 years ago and I work 10 to 12 hours a day from Monday to Saturday. Something around that. Maybe not every day, but something around that.

What about building in public? when, since you started sharing, when did you feel like you started having these inbound queries, inquiries?

It's really hard to tell because it's not like a hard line, it's like a gradient, But I think that even after a few months they started to show. But as you go up and up and up, better deals start to come up, know? So it's a process.

Yeah, it's a process. Yeah. Cool.

But for example, for me, one of the biggest breakthroughs were after I wrote my book. So I had one of these big breakthroughs along my career.

Mm. So writing a book is also a good idea, potentially.

Yeah, but he's hard and he's a gamble. I did not, again, I did not have the expectations that my book would do so well. Like, I had zero expectations when I started writing it, like completely zero. I said, okay, if I sell 1000 copies, I will be extremely happy.

Right, okay. Mmm. Yeah, and your best bestseller.

Yeah, we sold like over, like around 20,000 copies.

Congratulations. Yeah, it's a very good book. Yeah, I can share the link of your book as well to, ⁓ you know, in the description for our audience as well. Yeah.

Thanks. Yeah, that will be awesome. we were actually, me and Naksim started working and discussing about the second book, which will be around agent engineering. Who would have seen that coming?

What? Okay. you have to write it now. Yeah, tell me about it. What's your plan when it's coming out?

Yeah. I cannot tell too much because we are still cooking the idea and we are not yet sure of many decisions. So we are actively cooking on the outline and everything. But basically it will be around agentic engineering and it will take a similar approach to the course that I just explained. Again, taking someone who has zero experience in the agentic engineering space and building up.

Okay. Okay.

⁓ the whole landscape, but around a project similar to ⁓ the first book, because I have seen that people really like this approach of like actually building one large project where we show how to connect all the dots on how to collect data and not just, okay, we pull this data set from whatever on how you actually collect data.

Mm.

how you actually clean it, use it to find specific models, to build the memory of your agent, to ship your agent, to build orchestration, to do AeVals and everything that you need to actually build agentic applications.

Yeah. Okay. So it's your course or you could wait for the book to come out as well. So both seems to be change. Yeah. I'm curious. I know that last time we talked about you chose to stay. I think you are moving to Portugal, right? And versus you could have come to San Francisco and you chose Portugal.

Yeah, yeah, that's the gist.

Why is your startup is in SF? Why aren't you coming here?

Yeah. Yeah, well, actually, I like Europe. ⁓ I like the lifestyle from Europe. ⁓ So that's one big, big, big choice. It's also very hard to move continents ⁓ with everything. Also in America, you need a visa, which makes everything more complex. And I like my remote style.

Okay, I don't blame you. Mm-hmm. Mm. Yeah.

work setup and if I'm not constrained to be there, to be like in the office and all of that, personally I like more like to work by the sea, ⁓ by the ocean, like in a more relaxed environment. I don't like big cities ⁓ and agitation and things like that. It's not for me. So it's probably like a personal

Mm. I just. ⁓ okay.

lifestyle that I chose. But at the same time, Portugal is very connected to everything, like times and wise, it's closer to the US, it's closer to Europe. So I can very easily like run a remote business from there. Now in Romania is pretty hard because like it's the eastern, most eastern part of Europe mostly. So it's trickier.

Mm. Right.

And also in Portugal it is cheaper to live, lower taxes. it's like people who are familiar to like digital nomad ⁓ type of choices.

Right. They're all moving to Portugal. I have a family friend who actually took the whole family from Silicon Valley to Portugal. yeah, yeah, Yeah. He's they're very happy about it. Yeah. So I'm thinking maybe that's the way to go. But here's the thing, right? A lot of people are thinking about Silicon Valley. Do feel you miss out on not being in Silicon Valley or does it really matter?

really? How does he feel? Yeah, I do feel that I miss out and I think that when I made this choice, like I need to be very conscious about it. Like when I do have this feeling, I need to like understand and be honest with myself. It's okay. I choose this. It's a trade off. It's a bet. You you cannot have it all. And the thing is that again, for my type of like work, And lifestyle, I thought a lot and I don't think that I needed. Like there are of course some pros of being there, like being more connected to the community, having a lot more easier access to all kinds of conferences and things like that. Otherwise, like for example, the cost of living is higher, the taxes are higher, everything is higher, which means I will need to reflect that in all my prices and everything, which will make my business less competitive at the same time. So I ended their trade-offs. And again, me personally, I'm not that of a people person that much, like in the sense that I'm an introvert and I get very consumed by when going into...

meeting people.

Yeah, we meet people often so I will most personally... No, I I love discussing but now I need three times of recovery, three days of recovery. So, yeah.

I hope this is okay. my goodness. I will treat you coffee. I will treat you coffee in Pottugilco or something. It's a promise. promise. Lunch or dinner or coffee. I will go visit you when you move there. Awesome. Yeah. One last question. mean, you know, one thing I keep thinking these days, I mean, I grew up, I think a lot of us grew up, you know, being told to think about long-term.

Yeah, I knew that. So.

Right. ⁓ But technology and AI are moving so fast now. Do you feel is there even a long term anymore?

Well, it depends a lot on how you frame long term. I also like to think long term. I always think long term, but it's hard to know what long term is. So maybe five years is long term thinking now. And I like this way of thinking that it's okay to have a vision, a direction on where you want to reach. and like a few key bigger points that you want to reach, but to reach to that point you need to be extremely flexible. I think Jeff Bezos ⁓ said this like, you need this trade-off between, okay, you have this set of values and you want to reach to a specific point, like that's the long-term plan, but to reach there, things change so often that you need to be super extremely flexible. Like, I never imagined I would do content creation two years ago or three years ago or something like that. I see you just need to adapt.

Yeah, part of it is, know, how I grew up, know, my parents would tell me that, you know, your goal is to go to the best universities and then go work for Google or one of those big companies. I think it's really hard to imagine. Like, I don't even know what younger people today is going to think about when they graduate, what they're going to do. Right.

Yeah, yeah, it's harder to think about it,

That's the heart. What should we do? Do you have any advice?

⁓

Take your course and at least understand AI agents.

Yeah, no, I'm not the type of guy and I have to be honest I would like to say that I know why what the answer is I don't know like but personally I'm enjoying the ride at the moment like I love what it's happening at the moment and I think that AI can like make you like a superhuman in so many ways, especially if you're more into like a more entrepreneurial creative way where you want like to do a lot of stuff on your own. But I think that the most, the biggest advice I can give is just learn how to use AI first and own a domain, like be a...

Mm.

extremely good in a specific domain rather than know how to code and things like this. I think that's know how to solve a particular domain, whatever that is, medicine, finance, e-commerce, education. For me, it's education right now with AI. And if you go down, down the route, it will be very easy to adapt. ⁓

Mm.

That's

Yeah.

how I see it. It's hard to predict how things will look like. But if you have the domain, that's a thing that's extremely valuable and most you will still need like, humans will still need doctors, like whatever happens. But how doctors will look like and how hospitals will look like with that, don't know that.

We don't know. Yeah, we don't know. We'll see. Yeah. Yeah. Thank you, Paul. I mean, I really enjoy chatting with you. I mean, I think you are one of the rare engineers who build your, you build your AI career from somewhere. As you said, ⁓ the scene was zero, right? No AI scene over there in Romania. And you wrote your best selling book. ⁓

Hmm.

And you, I feel like you're, you know, emerging yourself in the trenches of building things, um, and sharing that knowledge in your course. So I really, it takes a lot of like energy to build and share. Right. And especially if you say you're introverted, right. I got to believe that must be scary when you first start posting. I'm like, I was terrible. Right. Like I have no idea. You were.

Yeah. Yeah, it was terrible. I was super anxious.

You made it!

I still am when I have to do a video, things like this. I need 50 minutes of preparation.

Yeah, yeah, yeah, I know, I know how that feels. And I feel that as well. And thanks for sharing that. It's really important for everybody, like all the engineers to reflect on like what they're doing, what they can, what more they can do, where they should go. And thinking of like overcoming these mental barriers of moving forward, right?

Yeah, yeah. Yeah, but like to conclude, I'm not demonizing like anything like even Big Tech or Cisco or... Yeah, yeah, of course, but nothing is that it's just you need to take decisions on what actually aligns with your true self. If that's not like too vague or whatever, but...

Okay, we're not saying anything's dead. Big Tech is not dead. Rob is not dead. MCP is not dead. Nothing is dead. Right.

Some people might like one thing, some people might like other things.

Right.

You shouldn't take like, the path that your parents told you, like, go to the best university, go to big, go and work at Google and that's the path.

This is the opportunity. Now is the opportunity to be yourself, maybe.

Yeah, let's hope so. ⁓

Right? Let's hope so. Yeah. Thank you so much.

Thank you. Thank you for inviting me.

Yeah, don't hand up, please.