...to learn ancient things.
The thing is, after the reorganization, it seemed the right project for me wasn't coming along. So, after the training in big data and the mailing phase, an opportunity arose in a document management project... (I confess I had to look up the name because I'd forgotten it)... Projectwise, by Bentley. So I jumped right in, and I made it to the fourth day of the training, thankfully they recorded the previous three. And you might be wondering, "Did you know about that document management system?" Well, it was the first time I'd heard of it (and hey, there are others I haven't worked with but they sounded familiar), but this one... nothing at all! :D
Thankfully, once you've seen one, you've seen them all, and this one had an API that could be accessed from PowerShell, so why not learn a new scripting language and access the API with it? And hey, it's not that complicated, and it's helped me better understand how the document management system works and automate tasks I used to do manually. So in the end, everything turned out well, and I learned something new, even if it wasn't what I expected (copilot completed this part :D).
The thing is, it was a typical situation where they needed someone to quickly take the reins of something that another team was going to stop maintaining, and it was the kind of product nobody knew, so... why couldn't I learn it? And well, the weeks went by, the team grew, they found someone in Portugal who was already working for the client in other areas and who could lead a team for the longer term, and finally I was able to step back.
Although some curious situations arose, this maintenance work had its charm. But hey, I managed to get out of that situation, which wasn't exactly a career path.
...it came to an end.
The next project, which I'll tell you about now, was one where, almost the first thing they asked me was if I was interested in changing divisions at Accenture, from Song to Technology. At first, I said I didn't mind and started working, but then, it turned out my mentor left the company, and she told me that if I had the opportunity to change divisions, I should take it, that there were more opportunities for growth in Technology, so... why not? I've been there for a few months, but hey, I wasn't in a hurry, so in the end, I was able to make the change.
...over a low heat.
It was October 2024. The project used Spring Boot 2.7, Java 11, Liquibase, and Oracle, with Angular 11 as the front end. A bit old, but not too bad. The thing is, I was used to technologies that were over 20 years old, and I wasn't going to complain about using something that was only 10 years old.
The thing is, one of the first things they complained about was the typical coupling between front and back end, that the front end had to wait for the back end to finish. Things that were really common 20 years ago, but fortunately things have evolved, and in one of the latest projects they were using OpenAPI for exactly that, so I suggested it, why not? If there's a way to start developing front and back end in parallel, after agreeing on an interface, why not take advantage of it?
But the resistance to change was noticeable at different levels of the project. I've always liked open-source solutions, because there's always someone before you who's had to solve a problem, and most of the time (for various reasons), they tend to share the results of that research with the world. The thing is, I set up a Docker Compose (that devilish invention), which ran several things, including Prims, a rather interesting mock server, and Killgrave. But it turned out that the "excuse" of security prevented the wider use of these tools. Then, curiously, months later, they started using them in other projects. And then nobody remembered, but oh well.
Another topic that came later was "Metadata." The question here was... trying to improve the automation of tedious tasks... We started with the functional definition, from there we looked for ways to automate the tedious parts, and then, integration with the development cycle, so that it would update automatically and could be used on a daily basis. But anyway, that's another story... one that will remain in my repertoire of learned things.
And well, the months went by and things stabilized. Trips came and went; that's the nature of the SAFE methodology. Every 10 weeks you have to synchronize and evaluate. And well, everything can be improved, but the dialogue was fluid, and in some of those synchronizations, I began to sense that the pace was slower than I had imagined, and that change was slower than I would have liked. And that was reflected in the people involved, and just as the projects were shaped, so too were the people involved. And if you wanted to flow, you had to too, although there was always some detour to try your luck.
...there are a few other things that could be mentioned, but I'll just highlight the arrival of Gemini. The truth is, we only had it at the chat level, without integration with any IDE, but even so, it was quite useful. One curious thing that came out of it was a diff for Liquibase, which was created from a diagram made in DrawIO and converted into a Liquibase code. And hey, it wasn't half bad...
...code and much more.
And although we've been immersed in a wave of generative AI for almost three years now, things are changing at an incredible pace every day. And back in the fall of 2025, I decided it was time to jump on the bandwagon definitively and without excuses. My goal was to get to the heart of AI; I didn't want to stay on the surface as a mere user, and initially, I didn't want to become an ML researcher, but I did want to learn the basics to move to the intermediate level, the metal, the code, the unseen inner workings, and if I could do it with Go, all the better; as I've already mentioned, it was my second language.
So I took advantage of the situation and set up a local server on a tight budget (and luckily, it was just before the RAM crisis, brought on by the voracious appetite of AI). The budget allowed for an Intel Core i5 Ultra 2nd generation, a 500GB PCIe Gen 4 NVMe SSD, only 16GB of DDR5 RAM, and a dedicated 12GB Intel Arc B560M graphics card. With this, I could start tinkering without having to rely on anything external.
The first thing I tried on that server was setting up a KVM and downloading a HomeAssistance image. They provide images for direct use. Under that ecosystem, I set up the connection with an LLM, and then with their agents, Whisper and Piper (interestingly, using a special protocol for HA)... Here I played around with various system prompts, including one to help my daughters with their homework.
In those first attempts, I don't think I had bought the GPU yet, so I was using the NPU, or even just the CPU. I was starting to see the Intel and OpenVINO ecosystem, understand how model quantization worked, and how the model server got the most out of the Intel hardware, and hey, it ran more or less smoothly. I also did some testing with Ollama, and even ventured into a PR.
The thing is, before starting all this, what I had been exploring was LocalAI, and the whole ecosystem that had been built around it for RAG. One day, while talking to some relatives about a video, I decided to look for a tool that could summarize the YouTube video, and there they were. That led me to try to build something similar (and I haven't quite managed to get it 100% working yet, but oh well, I'll give it another try later). Once I had LLM up and running, I abandoned LocalAI because Intel support wasn't very good (at least in my case), and I finished fine-tuning the model server. I also did some tests with OpenWebUI, which is actually a pretty complete system, even with RAG.
And after watching NYU's 2020 ML course years ago, I decided to take it seriously and leverage my ecosystem to help me learn. I set up a Juniper Notebook for Python and Go (although I haven't quite figured out importing external libraries yet). And I got down to work learning the basics (and I'm still at it, as I write these lines). I also wanted to explore the Go options within the ecosystem. While it wasn't the native language for ML, there seemed to be some interesting things. I've been diving headfirst into Hugot...
And here at this point is where I returned to my reading, the books I already had: Kahneman, Foer, Feldman, and the new ones to come... Clear, Ericsson... even others on how writing helps, and why I can't draw well. A few months ago I managed to finish a personal project I'd had open for a long time: the publication of a children's story.
And finally, back to the talks, asynchronous in this case (I didn't invent that), but resuming groups I'd tried to set up a while ago, because why not include an agent with visionary skills? If you take a look at my LinkedIn, you'll see what I mean.