AI & Technology

Over the past few years, artificial intelligence has become a stable part of my work through automation, software development and information analysis.

Once it moves from prototype to production, however, it becomes part of a wider system involving APIs, databases, permissions, costs and infrastructure.

That transition is what interests me most: understanding where AI is genuinely useful and how to keep its behaviour under control.

Applied AI and automation

Before choosing a model, I try to understand where time is being lost, which activities are repetitive and which decisions can be automated without introducing unnecessary risk.

For me, the most immediate value is in automation.

Many processes still consist of manual steps: reading a request, extracting information, classifying it or moving it between systems. AI makes it possible to automate some of these tasks even when the input comes from text, documents or emails that are not perfectly structured.

AI and software development

I also use AI in software development to analyse code, build prototypes, identify problems and speed up repetitive activities.

For me, the main advantage is reducing the time between an idea and a test. I can build a prototype, validate it and discard it quickly if the approach does not work.

Generated code does not receive special treatment: dependencies, configurations and changes need to be reviewed just like any other code.

Retrieval, data and knowledge

Another area I work on is access to information distributed across documents, databases, wikis, emails and different applications.

Techniques such as RAG — Retrieval-Augmented Generation make it possible to connect generative models to specific information sources.

Before generating an answer, the model can retrieve up-to-date information from controlled sources rather than relying solely on what it learned during training.

This is useful for internal search, documentation and support, but immediately introduces issues involving permissions, confidential data and source traceability: the system needs to know what it is allowed to read and what it is allowed to show to each user.

Security, agents and human control

Prompt injection, data leakage and credential management become concrete problems when a model has access to internal data or tools.

A system that can only generate text has limited capabilities.

A system that can read databases, call APIs, modify files or perform operations has real privileges.

At that point, controlling what the system is allowed to do matters just as much as the quality of its answers.

An agent can receive an objective, analyse a context, choose tools and perform a sequence of actions to achieve a result.

That flexibility requires clear permissions, logging, operational limits and the ability to stop the process. This is what I mean by controlled autonomy.

A system may classify information autonomously; if it needs to spend money, publish something, access sensitive data or modify a system, I prefer to require confirmation.

From demo to production

In production, APIs, databases, authentication, monitoring and logging matter at least as much as the model itself.

My experience with servers, cloud infrastructure and Google Cloud Platform leads me to look at artificial intelligence from an operational perspective as well.

A working demo is only the beginning. In production, costs, errors, capacity limits and availability issues emerge that are often invisible in a prototype.

Cost also quickly becomes a constraint: API calls, tokens, processing and storage all increase with usage.

Sometimes a smaller model is enough; in other cases the problem is better solved without AI, using deterministic software.

AI Project Management

I have also studied these topics formally through an AI Project Manager programme at Cefriel.

Project management becomes particularly useful when an AI prototype needs to become part of an organisation’s real processes.

Choosing the model is only one part of the problem: you also need to determine which data to use, how results will be measured, who is responsible and how the system will fit into everyday work.

The biggest difficulties often appear when the system has to integrate with existing processes and be used by real people.

Experimentation

As with SEO, I continue to approach AI through experimentation.

The speed at which models, tools and methodologies change makes it difficult to rely exclusively on theoretical knowledge.

I prefer to build small prototypes and measure what happens. If an experiment fails early, it has often still saved time that would otherwise have been spent pursuing the wrong approach.