This edition of SouJava Sessions brought a perspective very close to enterprise reality:
companies have large volumes of documents, repositories, internal pages, PDFs, knowledge bases,
policies, READMEs, intranets, and content spread across many different places. The challenge is
not only having information. The challenge is finding, organizing, validating, and using that
information in a useful way.

As part of the SouJava JUG, SouJava Brasília continues to create spaces that
bring the community closer to the topics shaping the future of software development:
artificial intelligence, cloud native, Java, open source, career development, and enterprise
solution architecture.

A journey built with Java, open source, and cloud

Paulo began by sharing part of his own journey. Born and raised in Brasília, he started his
professional career around the year 2000, working with Java before many of the frameworks that
later became common in the market were widely adopted.

Throughout his career, he went through different phases of enterprise Java evolution:
legacy applications, modernization, middleware, JBoss, consulting, cloud projects, DevOps,
GitOps, MLOps, Kubernetes, OpenShift, and, more recently, artificial intelligence projects.

With more than a decade working at Red Hat in the United Kingdom, Paulo brought to the community
a valuable perspective: the view of someone who deals daily with real problems from large
companies and needs to connect technology, architecture, security, data, and operations.

The problem: knowledge scattered across companies

One of the main points of the presentation was the problem of scattered corporate knowledge.
In many companies, important information is distributed across repositories on GitHub, GitLab,
or Bitbucket, Confluence pages, internal documents, PDFs, intranet pages, HR policies, technical
manuals, and old documentation.

This scenario creates several challenges. Some information may be duplicated. Some may be outdated.
Some pages may not have a clear owner. Certain documents may be difficult to find. And new team
members may depend on links, bookmarks, old messages, or the memory of more experienced people
to understand how things work.

This is where artificial intelligence can create value. Not as a magic tool, but as a way to
organize, query, and make accessible the knowledge that the company already has.

The challenge of enterprise AI is not only answering questions. It is transforming scattered
knowledge into a reliable, organized, and useful foundation for the business.

AI assistants with corporate data

Paulo explained that many companies want to create intelligent assistants capable of answering
questions based on internal data. However, this kind of solution must take security, privacy,
governance, cost, and model lifecycle control into account.

In sectors such as banking, financial services, and large corporations, data is often sensitive,
and environments may be restricted or even disconnected from the internet. Because of that, simply
sending internal documents to a public AI service is not always an option.

In these scenarios, training or adapting models with customized data inside the company’s own
infrastructure becomes an important alternative. The organization keeps control over the data,
the model, the execution environment, and the update process.

Red Hat Enterprise Linux AI and InstructLab

The presentation used Red Hat Enterprise Linux AI and
InstructLab as its foundation. Paulo explained that InstructLab works as the
open source project related to this approach, while the enterprise version goes through Red Hat’s
validation, testing, bug fixing, and support process.

The idea is to reduce the complexity and cost of creating AI models or assistants adjusted to
the company’s context. Instead of training everything from scratch, the proposal is to start from
a base model, add specific knowledge, and generate a new model that is more aligned with the
business need.

The demonstrated flow involved a base model, synthetic data generation, taxonomy files,
fine-tuning, validation, and running a model server to test the result with a chatbot.

This approach is especially relevant because it presents a practical alternative for companies
that want to adopt AI without giving up control, governance, and security.

Docling: preparing documents for AI

A central point of the demonstration was the use of Docling, a library capable
of processing different document types, such as PDFs, HTML pages, DOCX files, and other formats,
converting that content into structures that are more useful in an AI pipeline.

In the demonstration, Paulo showed how Docling can be used to convert documents into
Markdown files. This step matters because it helps organize the content and
prepare the information for later phases, such as taxonomy file creation and synthetic data generation.

In a real-world scenario, this collection process could involve Confluence pages, repositories,
internal documentation, knowledge bases, and other corporate content. The goal is to transform
dispersed information into a format that can be processed, validated, and used by AI models.

Docling Java, Quarkus, and Java’s place in AI

Even while discussing tools that are strongly present in the Python ecosystem, Paulo made it clear
that Java also has a place in this conversation.

He presented alternatives such as Docling Java and an extension with
Quarkus, showing possible paths to integrate this kind of processing into the
Java ecosystem. The idea is to allow Java applications to connect to conversion servers, send
documents, receive processed responses, and integrate this flow into real applications.

Quarkus also appeared as a modern alternative for cloud native Java applications, focused on
containers, GraalVM, fast startup, low resource consumption, and integration with environments
such as Kubernetes and OpenShift.

This part of the session was important because it reinforced that Java developers do not need
to stay outside the AI evolution. There are libraries, APIs, frameworks, and integrations that
allow Java to be part of modern data, AI, cloud native, and MLOps flows.

Taxonomy files and synthetic data generation

After preparing the documents, Paulo explained the role of taxonomy files.
These files work as a kind of knowledge guide for the model, usually structured with questions
and answers based on the content that the company wants to teach or reinforce.

In practice, extracted and organized content can be transformed into a question-and-answer file.
This file then serves as the basis for synthetic data generation and for the model alignment process.

This flow helps reduce the path needed for training. Instead of relying on a massive amount of
raw data, the company organizes relevant knowledge in a more directed format, allowing the model
to learn with greater focus on that specific domain.

Fine-tuning and customized models

Another important technical point was the fine-tuning process. Paulo demonstrated
how, starting from a base model and synthetic data generated from the taxonomy file, it is possible
to create a new model more aligned with the organization’s knowledge.

During the demonstration, the flow used InstructLab, a base model, taxonomy validation, synthetic
data generation, training, and the initialization of a server to test the adjusted model.

This process shows an important difference between simply using a generic AI tool and building an
enterprise AI solution. In a corporate context, it is not enough for a model to answer in a general
way. It needs to answer based on reliable, updated, and controlled knowledge.

RAG, vector databases, and knowledge updates

The presentation also touched on concepts related to RAG and vector databases.
In an enterprise architecture, documents can be processed, converted, stored in a vector database,
and queried through an information retrieval system.

This kind of architecture allows an AI assistant to use more updated and specific data instead of
relying only on the general knowledge of the model.

The combination of RAG, taxonomies, fine-tuning, and data control creates a more robust foundation
for enterprise assistants. Each approach has its role: retrieving information, organizing knowledge,
aligning responses, and keeping the model useful for the company’s context.

OpenShift AI and MLOps

After showing how to prepare data and adjust models, Paulo moved to an essential question:
how do we take this to production?

This is where Red Hat OpenShift AI comes in, presented as a platform focused on
MLOps, pipelines, notebooks, model training, model servers, deployment, and enterprise operations.

Paulo explained how platforms like this help organize resources such as GPU, CPU, memory, libraries,
workbenches, Jupyter notebooks, training pipelines, and model publishing.

This perspective is important because, in real companies, AI does not end with a local demonstration.
It is necessary to think about environments, scalability, security, versioning, operations, monitoring,
and governance.

Enterprise AI requires architecture

One of the main messages of the session was that enterprise AI requires architecture. It is not enough
to have a model, a chatbot, or an API. It is necessary to think about the entire path: where the data
is, how it is collected, how it is converted, how it is validated, how it enters the model, how it is
versioned, how it is served, and how it is monitored.

This perspective brings AI closer to traditional software engineering. AI systems also need lifecycle
management, quality, security, automation, documentation, observability, and an evolution strategy.

For Java developers, architects, and cloud professionals, this is an important opportunity. AI needs
people who understand systems, integration, infrastructure, data, and operations.

International career and the new developer profile

In the final part of the session, Paulo also shared reflections about career development, especially
for those who are starting out or want to pursue international opportunities.

One of the strongest points was the warning that AI may reduce space for very basic tasks, such as
simple CRUDs and repetitive forms. But that does not mean the end of programming.

On the contrary: programming remains essential, especially in more complex scenarios such as AI agents,
model training, automation, cloud, security, infrastructure, integration, MLOps, and system architecture.

Paulo highlighted that the international market values professionals capable of programming for cloud
environments, working with automation, understanding infrastructure, contributing to open source projects,
solving problems, and demonstrating knowledge through public repositories, personal projects, and community
participation.

The advice was clear: do not rely only on a résumé. Build evidence of what you can do. Publish code,
participate in projects, create labs, test tools, document what you learn, and show your ability to solve
problems.

Programming is not over, it became more demanding

The session also reinforced an important message for developers: programming will not simply disappear.
What changes is the kind of problem the market expects professionals to solve.

Basic tasks may become increasingly automated, but complex problems still require critical thinking,
architecture, technical depth, and the ability to investigate, test, integrate, and operate systems.

In a world with AI, developers need to become more complete. They need to understand code, data,
cloud, security, automation, business, and operations. AI can help, but it does not replace the
technical responsibility of those who build real solutions.

Final thoughts

SouJava Sessions #004 brought a deep conversation about artificial intelligence applied to the enterprise
world, with a practical view from someone working on real projects involving cloud, open source, DevOps,
MLOps, and modernization.

Paulo Menon showed that creating AI assistants for companies involves much more than connecting a chatbot
to documents. It involves preparing data, organizing knowledge, creating taxonomies, generating synthetic
data, adapting models, testing responses, operating environments, and thinking about production from the start.

For the Java community, the message was especially relevant: Java continues to have a place in modern
architectures, including AI flows, cloud native applications, Quarkus, integration with data tools, and
enterprise solutions.

Artificial intelligence is changing the way we build software, but it is also opening new opportunities
for those who understand engineering, architecture, data, cloud, and problem solving.

SouJava Brasília continues as a living movement: where knowledge, technology, and community come
together to shape better professionals and build more relevant solutions.