OK, great. Good evening, everyone. My name is Hugo. I am a software engineer.

And I just wrapped up after five years at CERN. And now I'm offering my service to companies who are developing AI workflow or agent tailored to their needs.

Tonight, I want to show you how to build an AI agent without the need of writing a single line of code. Because in 2025, now it's possible to build this agent without being a software engineer.

So first, let's come back on the principal milestone in terms of AI over the last years. So 13 years ago in 2012, we've seen the first model that has been trained using deep learning. So deep learning, it is the process of gathering a large amount of data and putting it into a model by training it with a large amount of GPUs.

This model, named AlexNet by ImageNet, has totally smashed its opponent in terms of the ability to detect features in an image. And 10 years later, ChatGPT is released by OpenAI. It runs on GPT 3.5 and It is the first large language model brought to the public.

It totally changed how humans interact with requesting data or information. It has appeared to a threat to Google, something that would not have been expected for such small companies. And three years later, here we are, AI is more and more democratized into a different process. It went multi-modal, which means it's not only able to interact with text, but also with images, audio, or even videos.

Just a few days ago, we've seen ChatGPT, they released a new image functionality with GPT-4.0 and it's constantly evolving. Where are we heading to? AGI, which is artificial general intelligence, which is AI that would be able to perform any kind of task, intellectual task by a human. In theory, it would even be able to run a company end to end. Personally, I think it will not be under the transformer architecture, but let's see in the future.

So now I want to explain what exactly is an AI agent, because we need to draw a line between an agent and a workflow. So a workflow, it is a sequence of events. For example, I will take a Google Sheet information. then I will run a large language model to request some kind of information over it, and then I will go to a next step.

But there, if we think about the workflow, the AI has one task and only one. It will not do multiple things. It's really controlled environments. What is an agent here? An agent is powered by a large language model. It has a memory. It has a variety of tools, an objective, and also an environment.

We can think the environment as a context. The AI has some information. It can pull information toward its objective. For example, it receives an email.

It is able to gather data in Google Sheet if required. Or it can read into your CRM. It can send you a message on Telegram if you want to implement human in the loop, or it can browse any kind of tool or different context provider. I heard questions about MCP, and this is, for example, something that is doing a lot of noise right now.

It's many softwares. They give, let's say, an API that is really for AI. So AI are now able to interact with other softwares and increase their ability to interact in a larger environment. How to build an AI agent without code.

Today I will present you three tools and I will even do a demo on the last one, NA10, but Michael already did it for me, but we will take more time to understand step by step how to do it. Each of the tools have their specificity. I believe that there are some that would be more interesting if you have very simple workflow. There are some that have a lot of integration.

When I say integration, it's the ability to communicate with another tool such as Gmail or HubSpot or Telegram. So let's start with Zapier. Zapier is by far the easiest tool to use.

It has more than 7,000 integration, which is by far the best at this job. And you can see it has also these cool copilot features to explain what you want to do, and it will build the automation. You just have to fine-tune it for your need. On the pricing, it has 100 euros for 10,000 ops.

One ops is one block here. One block equals one operation. So if there is five block in your workflow, it will cost four or five operation. And it has a free plans with a certain amount of operation.

The second tool I want to present to you is Make, which is a bit harder to use, but nothing really hard. It has more than 2,000 integration. It offers a free plan and 20 euros for 10,000 ops.

Make is nice if you want to bring more logic, if you want to build something a bit more specific to your needs. And the last one, which is my favorite, and not only me, I think, NA10. NA10 is a bit more technical.

I would say simple by surface. It's like if we take a programming language and we put it into a no-code shape. So it's blocks that are connected. The output of one block will be passed as the input of a second block, and each block has kind of its functionality.

It has more than 1,000 integration. And the pricing is quite interesting because while it doesn't offer a free trial, you have for 50 euro, 10,000 workflow execution. But a workflow execution, in this case, it is the full workflow.

So if it executes all this workflow, it counts for one execution. Where the previous tool, it would have counted for seven operations. Also, NL10, as Michael said, is open source, which means it's maintained by the community.

The code is available on GitHub. If you want, you can just take it and host it on the cloud. And everything about the pricing, you can forget it. It's yours.

Or you can implement it for your clients, for example.

So now I said NL10, let's build something. We are going to build a very simple agent, because we don't have much time. And we'll take the time to build it so I can explain everything.

The goal is simple. We get an email from Gmail. The AI will read the email, extract the characteristic.

And now we offer the AI two actions. Either it will create a task on Google task, and now on my phone it will vibrate because I will have some stuff to do. Or if it's about prospection, it will create a row in HubSpot trying to put as much enrichment of the data, like extracting the email of the person, the company, what kind of sector he's in. et cetera.

So let's try to keep it simple. Oh yeah, but then we'll have to do to do it like this. OK.

So I am in N8n, and I will create a new workflow. By the way, here you can see the URL of N8n is n8n.mywebsite, because I'm hosting it myself. And this is something super nice, because I keep control over my data in terms of the software. And as Michael said, if you want to have something that is really not letting any information going out of your server, it's possible with this stack.

OpenAI will read all your Gmail? No, no, no. OpenAI will have access to my Gmail information, of course. But I have the choice if I want to use any kind of model. It can be OpenAI, it can be Ultrapeak, or it can be something self-hosted, or something hosted on the cloud, for example. If I wanted something hosted on the cloud in Switzerland, I will have the possibility with this tool.

I will not let all my Gmail flow into OpenAI. I agree with you on that. But maybe some people, they will have no problem on that. The important is to have the possibility to choose and to offer your client to choose.

So for this automation, I need to add a first block that will be the trigger. So it can be triggered manually, or it can be triggered by an event.

And me, I want the event to be Gmail when I receive an email. So I will type the name of the integration I want to put here, Gmail.

And I see I have actions, but I don't want to do an action. I want something to trigger it. So when a message is received.

OK, here I am inside the configuration of my blog. So I can say on my Gmail account. It will fall every minute. OK, why not?

When a message is received. And this I will just remove it, because I don't want the body of my email to be simplified.

Now what I want, once I have the email, I want the AI to read the email, extract the characteristics, and then do one of the actions. So I will use the AI agent. block that is very powerful. And let's look at it first.

It looks like this. Because an AI engine, as I said, it's powered by an LLM. It has access to tools. It can optionally have a memory.

And it's up to us to define it and tailor it to our specific need. So here, we first need a chat model. Bad example, but I'm using OpenAI.

I will not leave it plugged to my email all the time, but it's connected to my OpenAI account. And now I can select a list of models. So these are all the models that are available on the OpenAI API.

And me, because I want to keep it cheap, I will just take the 4.0 mini. And that's it. There is the AI model.

Now, in terms of tool, I have two tools that I want to put. The first one is to create a task. So it will be the Google Task tool.

OK, so it will connect in my account. The tool description, so the AI know what is this tool. I will leave it automatic. I will allow it to create tasks.

The title of the task, either I can type the data to have it hard coded, or I can click on this magic button, which is letting the model define what this property should be. And obviously, I will leave it like this, because I want the AI to be able to take action on the task. I want the AI to choose what will be the title of the task. So I don't have to read the full email, but just browse my task list and read the title.

And I will add something else, let's say a note. And on the note, I have, once again, this magic button that I will press. So the task that will be created will have a title and a note.

So this is it for the task configuration. Or not? Oh, yeah. Here. My tasks. OK.

And I will add another tool, which is HubSpot. HubSpot, it's a CRM. I don't know if you know about it, but it's kind of like a big table that will contain the list of my contacts.

And here, it will, OK, again, description of the tool. I want to create or update contact. And the email of my contact, I will leave it automatic, like this.

And I can add some other properties, for example, the city or the first name. First name. Last name. And yeah, there is a lot of things. I will say company name. OK. And I put everything automatic.

All right, now there is the last thing I need to do for the simple agent to be working, is here the configuration. Because it receives input data, and I need to kind of connect it. Like say, OK, I want the subject of the email to be taken, and I want also the body of the email to be taken.

One thing that is very nice with NATN, it enables you to work step by step. It's a bit like programming language. You never write a full code in one shot. You do it step by step. You understand what data you have, how you manipulate it, and how you work with it.

So you have the ability to execute the previous node. So this is what I will do. I'm executing it, and it will just fetch my last email. It is like an invitation to an event that has been accepted. So I will just use it to pull interesting properties.

So I would just add a system message and the prompt, which are the data. that will be used to provide a custom enter. I will use the subject of the email here, you see, by taking it drag and drop, and the HTML. I will just do some formatting so it's nice. Here, body, I say the body of the email is like this. And you can preview it down what the email is, and the subject.

I like to give a bit more context to the AI, so I will also put the date. So the AI knows when it receives the email. Otherwise, it doesn't know. OK, great.

And now for the system message, I will have prepared it because I don't want to fully type, but we will read it. All right. Your goal, you are an AI agent built to classify email based on urgency and relevance.

Your tool, Google task. If the email receive an action, create a structured task with a clear title, description, priority level, and a due date. Assign it if possible. HubSpot API.

If the email is related to prospecting and only prospection, check if the contacts exist. If not, you can create a new record with all available details, including name, company, email, and relevant note. Do not use the HubSpot API if the email is not about prospection. So now we're good.

I can test the workflow. like this. Oh, and that was not what I expected. I will just send an email to myself, because I want to work with some materials.

So I will just open my email, put it there, and then write a new message to myself. And we can start with a prospection email. I'm sorry, it's in French. So basically, what happened here is there is one guy, Alexandre Martin, which sent me an email about prospection for some kind of solution, et cetera.

And he's also offering me here a moment to take appointments. So if I read this, I believe Wow, I need to read a lot of things. I believe I optionally have to answer him if I'm interested or not. And if I'm interested, it would be nice to answer before, let's say, the 1st of April. So this way, I will not miss the first moment he's available.

So I will just send this to myself. And whenever I receive it, I will be able to retest the workflow.

So on the previous execution, by the way, you can see here it's very visual how it worked, because the Gmail trigger block has worked in green. Then it went to the AI agent, where here it sends the information to OpenAI, to the model, which think, OK, I have these tools, and I have this information.

So the previous email, it was a real email. It was like someone send me an invite. When you receive an invite for an event, it sends you an email. There is nothing really to do about that.

And it picked no tools. So if there is nothing to do, it will just forget it and do nothing. Because what's the point of working if there is nothing to do? Now if I go to my email, I just want to make sure I receive it.

So I will refresh. And yes. OK. Great.

So now I will test the workflow again. And it receives it. So this was the prospecting email. Oh, it also created the task. OK. OK, right, it finishes.

So I can also click on this to understand what it did. So here it created a task, follow up with Alexandre Martin. Note, respond to Alexandre regarding availability for a call or meeting. OK, it tells me that maybe I should answer him.

And if I go to my Google Tasks, I will see I have Okay, it also created the event, nevermind. But it also created the follow-up with Jean Dupont from Logistic Express.

All right. And if I go in my contacts in HubSpot, oh, now I see two rows. So I see myself and I see Alexandre Martin. So it managed to get the information.

Okay, let's do another test. This time it is customer service.

I will send myself another email, which is a user that have an issue. He's not able to subscribe to the software. And when you're running a business, it's very important that your user are able to pay for your software.

Because otherwise, you're making no money. And there's no point doing a business if your user cannot subscribe.

So whoops, I sent it a bit too fast. But I will just go back here.

The user says here he has a bug. So when he reached the payment step, he has an error message that says impossible to proceed the transaction.

So at that moment, it's very important that we take an immediate action. And we understand the user request. We call him. We call the IT team. We fix our software. And then we say, OK, it's fixed. Now you're able to proceed with your user workflow.

OK, let's exit this block and run one more time the workflow. So it receives the email. Now it is thinking. And it failed. Oh, no. That's the curse of live demo.

Yeah, maybe there is no money. OK, I don't know why. Can we try again?

OK, I'm lucky it works. So it managed to create a task.

So the task is urgent, payment software bug resolution, and then a note that describes the issue the user had. You can know that the HubSpot tool is not circled in green, because it doesn't need to create a contact in HubSpot. When a user is submitting an issue, you should treat the user issue, not collecting his data. And in that sense, the agent made a smart choice.

It only used the tool it needs.

So this is something we built in really five minutes, but imagine now we give one more hour, all the functionality we can put. Maybe something to send me a message instead of creating a task for this kind of high priority.

There is much more to explore about this software. And if you want more, I encourage you to ask me questions.

Any questions for Hugo? Yes?

First, thanks a lot to both of you. I mean, because I've never really used Agent.

I've heard it many times before, and it's the best introduction I've had so far. But the first question comes to my mind when I see this stuff. I mean, is it really reliable enough that you can sleep well without looking at all the stuff yourself as well?

Because I'm really puzzled. I mean, if I wanted... For example, okay, the Gmail is, email is, of course, just an example, I guess.

But nonetheless, I mean, it's an example where you say, okay, this is what one could do. But you basically let the AI interpret your email and only take action, or basically you only read then afterwards the interpreted emails. So for that, it has to be very reliable because otherwise you would go back and read the original email as well.

So is it really so... Would you trust this stuff?

For this implementation, yes. But let's imagine we go a bit further and we let the AI optionally answer to the user. I would never do that because I would be so worried that even if it happened like 1% of the time, it would be too much, that the AI has an hallucination and answer something that is not relevant.

I don't want this. what usage of the AI you want and how impactful to your business it could be if the AI makes a mistake. So the AI is being better and better like every week and this we're seeing it just looking at the news.

Right now there are still some tasks you cannot do and this is why I was saying we are heading towards AGI but we're clearly not at AGI. So you must really think about what type of repetitive task you can improve in your workflow, and how the AI can be integrated, and how you can monitor the errors. And this is why, before interacting with a user, I would always implement human-in-the-loop validation.

Like, if I wanted to send an email to automatically enter the user, I would send myself a message with a draft of the email before entering, but I would never let it be autonomous. But still, you would let the AI decide of whether you contact this person that contact with you, whether it's important or not? If we take this example, it will be just creating tasks or adding rows in the CRM.

Yes, I will trust the AI for that. And do you know of companies that do that stuff?

So I think it was Google that said that more than 60% of their code is AI generated by the year 2024, so last year. So I believe, in that sense, they are already trusting the AI for coding. But they have human validation in the loop, which, in terms of code, is called a request.

Whenever you add code to the code base, you ask another developer to look at the code for you. So this is an example of human in the loop.

Here we go. Yes?

You both mentioned that there's a lot of pre-built integrations with these. If I was going to use this with mine, we have a custom-built software, and we have specific booking numbers for clients and everything.

How easy are these tools to... it's great when you're not using the pre-built integrations with Gmail or anything like that. When you have to call a specific API to find information So maybe you are using your own tool that you developed.

And in that sense, your tool, you may have an API if it's a software. This, you can totally interact with API just if I was writing code. This gets more technical.

We are getting out of, let's say, the nice blocks that are already available for drag and drop. It's more like an edge case, but it's possible with this tool. And do they have

primary tasks are they replying to emails and there is all sorts of automation. If I go back here and I look at what I can do, it's divided into category.

I have blocks ready for AI. I have action in the app. And then I can really browse the app, like Clockify, even Cloudflare.

And a lot of these, I don't even know what this is, but like Elasticsearch, for example. And you can do a lot of, there is a lot of integration. But then if you want to have something

tailored to your in-house software, you can write JavaScript in one block. And for example, the action to, for example, submit a form into your system or pulling data from your system, you can write a block with custom JavaScript that will just do this. And then the rest of the, let's say, workflow architecture, it will be with blocks.

And it allows to connect your system with other systems that are supported by the platform. Yep. Yeah.

Mm-hmm. Though I will not use this in production. If it's a large system, you will have issue scaling.

Like this, having 1,000 users, it would be OK. But then if you have 10,000 users, 100,000 users, then it's maybe better to do a real application using long chain, for example, because it would scale better. But if you need automation for a small company, that would be perfect.

And also, you can host it yourself. So there is no additional cost, except of the LLM. It's hosted on a server or on a local client?

This N810 is stored on my home server. At home, I have basically a computer that I recycle into a server. And I have a bunch of service on it, including my website and N810.

So I don't pay for the software, basically. So before if you wanted to build agents anything it's out in the market for if more than a few years now so it was already possible to build it but the main problem is the cost of the token because whenever you put AI in a product you will have like a cost because you need to pay for the API.

Self-hosting models, it's really painful. So except for really custom case, I think it's better to just ask, pay for a service. The cost of token is high, but it's decreasing exponentially and it's enabling you to do more and more automation.

So you are playing kind of in the same game as big companies. If they implement something that is not worth it, there is not a real point doing it. So just to expand on that point, this nice little agent here, I receive a bunch of emails and I'm going to use the agent to summarize them for me and maybe classify them, something like that.

What does that cost in terms of tokens, for example? It depends on the length of the email. It depends on the model you use.

Here I use GPT-4. A couple of pages of email and a few hundreds per day. I think 10 e-mails, it will- person who receives e-mail and reads it every now and again and does stuff, or I'm going to put in place an agent and then pay the costs.

Yeah. How do I do that? How do I do that calculation?

So for example, here, I'm using the OpenAI chat model, and I can go on the OpenAI website, and I will have the comprehensive section about about the API platform and how much it costs, everything I do. And here it should be a couple of cents for a few emails.

And then if it's scaled up on, OK, I will log in. And we will see what was the consumption. I think it will be a few cents for the few emails we did.

But in one year, it will be maybe 10 times less. There is a lot of investments on data centers all around the globe. And this contributes to lower the cost of AI.

In another sense, the AI companies, they are building their models to be more and more performance and to cost less token. So for example, this is why when DeepSeq released, it was very good because the model was as good as GPT-4.0, but it cost 10 times less. So I think it's not pricing.

Where is it? in the dashboard and in the usage. And we see today I did here.

So it's less than $0.01, because obviously we didn't use it a lot. And I'm really playing a lot with GPT-4.0, and it's only $0.26 for March.

I think it's really depending on your use case. Because if you receive big emails, it could totally change compared to small emails. But it really depends on how many emails you receive and how many tokens there is to analyze.

And I guess then, for using this, the other question is the data. Where is the data going? In the example that you have here, or maybe if you're in the example where you have the one hosted locally, is all the data, the content of that email, is that staying local on your machine, on your server, or is some part of it going out to...

Now it's going into OpenAI in the US, because they don't have servers in Europe. So for prediction, this is many companies, they don't want to trust servers that are hosted outside of Europe. But if I want to change this, I could go here in the model block.

And I could select, I will just remove this one, and add another model. And as Michael said, you can use for hosted models, or you can use It's a bit confusing, but basically OpenAI, they provide an API.

And many AI companies, what they do is they say, OK, we provide the same API, but it's our AI running on our servers. So you can use the OpenAI block. But saying, for example, if I have some cloud provider in Geneva, I could simply use the one from Geneva.

And then the data will never go out of Switzerland. This is something totally possible, and there are providers, cloud providers, that provide you models in Geneva with high security. Infomaniac.

Thank you. All right. Okay.

Closing Remarks

Yeah, well, I figured. So if there's other questions, you can obviously, like, everyone's invited to stay. The pizza's on us.

I also have beers. And... Thank you so much Hugo and Michael for your presentation.