Introduction

Welcome, everybody. Thanks a lot for the organization.

My name is Juan, and I'm going to talk a bit about AI adoption, how can we leverage in average companies, and how can we use AI in order to get better and more efficient and also make all the company work seamless

rather than just a bunch of people using chatbots or ghost AI, right?

So I was asking a bit the crowd. More or less, 50 % of the cloud is technical. The other 50 % is not.

But how many of you have used CloudRooter GPT? I hope everyone. So please raise your hands.

Okay, yeah. Yeah, that's everyone. And which of you have used Codex or Cloud Code? Please raise your hands.

Okay, more people than the technical ones. This means that they're doing something good, right?

And Tropic, OpenAI, we're moving really, really fast and they're replying news every week, every month. We have new updates.

Maybe this is just me, but I find it super difficult to adapt.

Who I Am and What I Do

So yeah, a bit about me, I work in a BC deep tech fund, which for the ones who are not into the slang Lang is basically a firm that invests in startups and helps them grow, right?

And basically what I do is I run the AI strategy and AI operations.

So I basically empower the team to use more AI, better AI, but trying to reach the final objective that is getting there, right? Getting in a way more efficient, way more scalable system

Why AI Adoption Feels Overwhelming

they were able to build really nice things, right? So, as I was saying, the market's moving pretty fast. You probably, all of you have heard news, newspapers, like everyone's talking about AI, and there's so many things that we're talking about, but it looks like a train that everyone's missing, right?

So, I've been following GPT and Codex, And some months ago, people were saying, oh, Cloud is the best thing. You should integrate it in your company.

Then people were like, oh, but Codex is also good. Codex is catching up early. And before, they were talking about ChatGPT as their own tool. And this has been only this year, right? So it's a change that we're seeing all the time.

So what I mean with this is you can no longer expect your team to be as fast as a centropic team that is building.

Like, these releases are people releasing it, but mostly it's agents releasing it, right? So you have to, at least I am feeling like, well, maybe this is a bit too much, no? So I have to develop a strategy in the long term.

I cannot be like, oh, but CharGPT is better, Cloud is better, oh, what about Gemini? Gemini also integrates good with my email, so probably I should do Gmail. Gmail.

From Tool Chasing to a Long-Term Strategy

So I don't have a really solution for that, but what I tell my team is we have to build a scalable system. We have to look for a process. We don't have to look for AI system as a tool.

We have to build orchestrating, build the system that works and is able to get us further, right?

The Core Shift: It’s About Context and Process

So how do we do that, right? You're probably thinking, okay, yeah, beautiful. Yeah,

let's build something scalable. But how do we do it?

It's not about the chat. It's about context.

It's a game where you have to give context to your company, and you have to give culture, you have to get education, and you have to really push forward if you want to be, okay, I'm going to get to the age of AI.

Building In-House: Goals, Inputs, and Reusable Workflows

And how can you do it? You can get people like Jorge itself or companies that integrate it for you, but if you want to build it from in -house,

from in -house without your own resources, you have to have really structured goals, explicit inputs, persistent things that are able to use one model and the other model, right? So, at the end of the day, it's how do you want to see your workflows going, right?

Workspaces and AI as an Operating System

So, as I was saying, Anthropic and OpenAI are building workspaces. What's a workspace? space it's somewhere where you live where you interact where all the magic happens right all

the businesses processes are starting to get more integrated you can be in the construction side you can be in the consulting side you can see you can be in agrotech you can work in a in the most beautiful business in the world or in the boring company that no one wants to talk about, but all of them are getting asked the same question. Okay, what are you doing about the AI, right?

So this application such as Codex and Cloud Code allow you to operate as a system, right? So you have to see, okay, where should I go or what should I build, no? That's what you're probably thinking, no? If I want to do something, okay, what should I do? do.

Demo: Agent-Orchestrated Work in Practice

So I'm going to do a small demo of how I run and how do I use the tools that I normally interact with.

What “Agents” Mean and Why They Matter

I built an infrastructure in order to have agents. I don't know if you are familiar with agents. Please raise your hand if you've worked with agents.

Okay, okay, Okay, okay, okay, perfect, perfect.

So I basically built agent orchestrator for my film. So basically I lead the product, but other people are leading, for example, investments, or other people are running operations or finance, right?

So I made an example by choosing the finance team because it's something more vertical and that everyone has, right?

Example Use Case: A Finance Expenses App

So we built an application, right, for expenses. So it's just basically you scan via QR code and via OCR, so basically you get an invoice and you're going to be like, oh, this is an expense I did, this is this day, and the administration teams approves it or rejects it, right? And it's a really, really nice product.

And this was built by one of our colleagues who works in Operation Analyst, and she basically bytecoded all the app, right?

But also, in a team, you have to be like... Oh, shit. Perfect. I forgot the Wi -Fi.

While my computer connects, I'm going to show you. you.

Deployment Readiness: Security and Compliance Checks

Basically, what I did is SheBipe coded the whole app, and I was like, okay, YouBipe coded it, it works, now we have to deploy it in a secure system, because that's one of the issues, you know, someone codes something, it works, it shows to the team, we start to implement it, and then someone is like, hey, have we thought about security, right?

Or have we thought about compliance, or where is that getting sore, right? So it's a question

you have to ask yourself and be like okay we we have to get there so i run a five agents right in order to to check the expenses up right so i'm gonna show you the app right so we we i first before everything i integrated the authenticator via microsoft in order to have like secure um

So, basically, what you do here, it's like you get your expenses, you see where you spend this. I went to an Italian restaurant and I talked with these people, right?

And then you have the administration portal where you can see, okay, oh, they've approved both of my expenses, no?

And she built everything without knowing a single line of code. She never coded before. for.

And we're seeing like this, that a lot of people are building things. And we have to, we have to see what are they building?

How are they building it, but the things that we have to give us product leaders or as the people, if you're here, maybe you have some sort of interest in AI. If you want to

bring AI to your team, the most important you have to give them is a framework, right? It's a place where they feel safe, and they can build on.

Creating a Safe Sandbox for Iteration

It's giving them the right tools to build, right? So, how do you do that? Iterating. For me, it's been the learning processes.

It's been a lot of, oh, wait, I haven't thought about this, or I haven't thought about this. So, I built like a sandbox where everyone can test things, right?

I was testing, I don't know if any of you have tested the codex expect, but basically you can call an agent that it's a pet, and basically it's going to pop up and it's going to be the representation of my agents and it's going to give me notifications on when is this working.

Using Shared Context Across Tools with Markdown Files

In the meantime, while it charges, have you seen this prompt? right? It's a pretty big prompt, right? But I'm using CloudMD.

Why is that? I'm in OpenAI app, right? Okay, for those who probably don't know, OpenAI has Codex, Cloud has Cloud Code,

right? So, I'm using my system, which I built on Cloud, and I'm using it in Codex, right? But can you do that?

Yeah, you can if you've built a markdown file, which is basically a text file that you integrated in your laptop, no?

So I basically said, OK, can you check the vulnerabilities of my app using Cloud MD as your main brain? And yeah, basically, it detected. It's in Spanish, so probably better English.

So it said, OK, wait, wait, wait. You have a lot of critical things, right? You have unauthenticated generated database API, right? What does it mean?

That basically anyone that has access to the app can auto -accept themselves or generate invitations or see the expenses of the rest of the team. That's probably something we have to fix.

Or unauthenticated Azure Blob, right? The endpoints, that's a storage thing think that 50 % of you are not digging into this, or authenticated users, right? Or authenticated proxy can burn API credits, right?

So these things are helping me detect what this endpoint or what this system was doing and helped me diagnose what was happening. So I said, okay, as you say, we're like, okay, perfect.

Automating Fixes: Vulnerability Detection and Remediation

Turn it on and basically, hey, can you with a goal make the first fix on the expenses app? Tell me first what it's going to do and spawn an agent to explain where you start deploying it.

So there's a lot of functionalities, right? Like goal, right? That it's basically going to deploy a bunch of agents that are going to act like a system

and are going to help me deploy. Okay, perfect. This is a direct. Sometimes it just doesn't work.

Perfect. Perfect. Right now, I'm using my Cloth files, like, when you use Cloth, it generates a text file, and when you, yeah, okay, not working.

When you generate a Cloth, and you interact with them, it generates a Cloth .md, right? Right? Everyone works with markdowns nowadays.

So, I basically said, like, okay, I'm going to run it and I continue talking because if I start talking and I don't run it, the demo is not going to work.

So, basically, what I did is I used the context that my cloud has in order to fit my codecs. Right? So, the first fix is close anonymous access.

So, So he has created an agent to revise in parallel the authorization and basically it's looking at the database, all the tables, so it's fixing the first issue that he was looking.

And he basically applied this change and it's importing directly. So, my pull request for the non -technical, basically search for authorization to post it under production, I can review it here and I can see what's it done.

So, it's going to fix all my cybersecurity issues, right?

Orchestration Pattern: Multiple Specialized Agents, One Conductor

While it works, I can also show a bit of the agent orchestration I'm normally doing, which is basically, for example, I have this to report to my manager.

So I have the different tools I have, and I orchestrate it all with the same system. So it will look something like this. right?

So it's the app context, the threat model, the code security, and all these agents are talking to themselves. And the orchestrator, which is the agent I called, is orchestrating everything.

So it creates like a symphony, right? So yeah, that's a bit of an example of how do I use.

Execution Discipline and Takeaways

But as I was saying in the beginning, in order to implement AI, it's not enthusiasm. We We have to have shared execution discipline, file systems that work with all the models,

and in order to create a culture that works with AI, we have to start with goals, share reusable specs, review with gates, and save all these workflows so everyone can use them and expose them and have weekly updates and people involved.

Because AI is going to fit everyone. one.

And every single person of us, we don't want to know how to code, but we want to excel in what we do as the best ones. So

if I want something for you to take as takeaway, it's the model is not the mode, the execution system is.

So go and build the execution system.

Conclusion

Thanks a lot.