In November 2025, I was writing Python code manually and learning how AI works. Since March 2026, we have been building NEURIXA on those foundations.
One of our starting questions was: how can we build an AI environment that can perform defined tasks locally, without relying on an external AI service?
From independent learning to building a system
During development, I initially used Codex CLI. Later, I incorporated Claude Code and Codex into my VS Code workflow. Drawing on my existing IT and information security knowledge, I used AI-assisted development to implement individual components and connect them into a working system.
My working methods evolved alongside the development tools. Implementing each function meant examining how the components fitted together and what conditions they needed to operate.
Which questions did we explore in practice?
- Which tasks can the available hardware support?
- How does the system move from a document to a useful answer?
- Which processing steps depend on external services?
- How do we verify the results?
- What is needed to make the operation repeatable?
Why build our own environment?
We wanted first-hand experience of the possibilities, limitations and operational responsibilities of AI systems. Building our own environment made data processing, component integration and output verification practical engineering tasks alongside model selection.
Local operation and offline capability need to be assessed separately. A component running on your own hardware does not mean that the entire workflow operates without external services. Dependencies must therefore be reviewed across the complete task.
A practical foundation for advisory work
These experiences also provide a practical foundation for NEURIXA’s advisory work. They help us consider implementation and maintenance requirements alongside business needs.
Suitability for enterprise deployment must be assessed against the specific task, data, security requirements and operational resources. Experience gained in our own environment provides a starting point, but it does not replace an assessment of the customer’s environment.
Which task would you test first in your own AI environment?
