AI approach
AI assists every project. It does not run them.
We use Claude Code and Codex every day. Agents when they help, not by default. A human reviews and tests every change.
How we work
01
A framework for each project
Before any code is written, we set up the context, the documentation, the harness and the restrictions that suit the type of project. We work out with you what is sensitive and what must not be exposed to AI.
02
AI-assisted development, always
We work in VS Code or another editor connected to AI tools. The development of your project’s code is always assisted by AI.
03
Agents when they are useful
Agentic programming is not systematic. We use agents when the task calls for them.
04
MCP where it helps
We use the Model Context Protocol (MCP) to speed up development, fetch information or trigger actions, depending on what the project needs.
05
Systematic code review
Every change to the source code is reviewed.
06
Systematic tests
Tests make sure that a change does not alter the behaviour of what already exists.
07
Git on every project
All source code is under version control, without exception.
08
Two backups, or your own tools
By default your project is backed up twice, in the cloud and locally. We can equally develop and store nothing on our tools and use yours instead (on-premise development).
What it changes for you
AI is completely transforming software development. What seemed impossible not long ago, for reasons of time, budget, complexity or maintenance, no longer has to be.
Building all of your software tools yourself is certainly not the best answer. But having hand-built components for what makes your company specific, and valuable, is now entirely possible at a reasonable cost.
140 h
to put the first version of a three-platform application into production, with AI involved from the specification onwards.
3 months → a few weeks
for a web application we built before AI and would build far faster today.
Years of waiting
ended for features long identified as needed but too expensive to build without AI.
Sovereignty of your data
You are entitled to a plain answer to a plain question: where do my code and my data go?
As of late 2026, we use Claude Code or Codex by default. Our prompts therefore go to the servers of Anthropic or OpenAI. These companies consequently have a very precise idea of what we are working on, and have access to the source code of the solutions we develop.
However, by strictly framing the development of each project, we do not by default give access to your data, nor of course to the credentials for your databases, your APIs or anything else that must stay confidential. These are stored securely, in a way that depends on the type of project, outside the source code that the AI can access.
When a dataset is needed to run tests and make sure the software works properly, we generally build a set of fake data. It is consistent with the context of the test, and it is never your real data.
We adapt this level of confidentiality and sovereignty to your needs. We already work with companies for whom this subject is paramount.
What the AI providers see
- Our prompts
- The source code of what we build
What they do not see by default
- Your real data
- Credentials for your databases and APIs
- Anything that must stay confidential
What can be adapted
- Development on your own tools and infrastructure
- No storage on our side
- Critical components written by hand, without AI
- A stricter framework, defined together
When AI has to stay out
Using Claude Code or Codex is not a dependency. NavArchSoft has been writing software since 2015, and most of the projects shown on this site were built before AI assistants, or without them. When a piece of code is critical, or must not be exposed to any AI provider, we take over: it is designed, written and reviewed by hand.
Where that line falls is decided at the start of the project: we work out with you what is sensitive, what can be entrusted to AI and what must never be shown to it.
The ANSSI recommendations
We rely on the security recommendations for generative AI systems published by ANSSI, the French national cybersecurity agency, and on the implementation guide that Campus Cyber derived from them. We can apply some or all of them, depending on what you need.
What we already do
- R2A risk analysis when the project is framed: what is sensitive, and what may be shown to AI
- R30Every piece of AI-generated code is reviewed by a human and tested
- R31Critical modules are written by hand, without AI
- R34Real data and credentials are never sent to an online AI tool
What we put in place on request
- R30Automated static security analysis (SAST) of generated code
- R30AI review by a different model from the one that generated the code
- R34Development with no online AI tool at all, on your own tools
- The rest of the guide when your project itself embeds generative AI: partitioning, logging, input and output filtering, trusted hosting
We claim no certification. These are practices, which we document project by project.
This website, for example
This site was built the same way: a written brief, documented context, AI-assisted development and human review. It uses no cookies and no trackers, and its only third-party script is the anti-spam check on the contact form.
A need, an idea, a spreadsheet that has outgrown itself?
Tell us about it in a few lines. We reply by email, and can follow up with a video call.