The problem I wanted to address
Companies can find practical uses for language models in document analysis and drafting, but the documents involved may contain sensitive information. The project began with a question: could a layer between a user and an external model help reduce the information exposed while preserving a useful response?
How the platform works
The platform masks sensitive business data before it is sent to an external AI model and restores the original values in the response. This creates a workflow in which document analysis and drafting can be explored with an additional privacy measure between the source material and the external service.
Thinking about the whole exchange
The useful part of the project is not only the transformation of the input. The response also has to remain understandable when the original values are restored. That makes the relationship between the input, the transformed text and the final result central to the product. It is a practical data-flow problem as much as an AI problem.
What the project changed for me
IAProTech gave me a more concrete way to think about product development. I had to start from a specific concern, define what the tool should do and consider the constraints of the workflow. It was an important step from experimenting with AI tools to trying to build a useful service around them.
Scope
The project explores a technical approach to data protection. Pseudonymization is one measure within a broader privacy and security process; it is not, on its own, a guarantee of legal compliance. I describe the project in those terms rather than treating a technical feature as a complete compliance solution.