Skip to content

All case studies

Technical documentation under the MDR: tools built with AI

Nine tools and a portal for regulatory affairs, built with Claude. They fill in templates, generate declarations of conformity and check references. Local in the browser, with no AI at runtime.

Result

Roughly between May and October 2026, nine tools and a portal came out of this: four tools are ready for use, five are in a test phase, and the portal is in use. I have not measured time savings for any of the tools. Two test runs with real data are documented. The declaration of conformity generator correctly split an item list with a three-digit number of items into three declarations, and the Word files open without a repair prompt. The tool for the technical documentation summary transferred several hundred reference lines from a real file into the right sections and marked unreleased documents exactly. A sample chapter was identical to the version made by hand.

Starting point

I work in regulatory affairs at a medical device manufacturer. Technical documentation, EU declarations of conformity, labelling and change management under the EU Medical Device Regulation (MDR) were largely maintained by hand there: copying data from ERP and DMS exports into Word and Excel templates, looking up references one document at a time, finding contradictions by comparing line by line. That takes time, and errors easily go unnoticed. There was no server for providing in-house tools centrally.

What was built

Nine tools and a portal that bundles them, each a single HTML file. I built them directly within the department, for the regulatory affairs team and for colleagues working on technical documentation. They all follow the same rules:

Creating

Checking

Deciding

Providing

How it came about

See all 5 steps
  1. Describing requirements in chat

    GitHub CopilotVS CodeClaude

    I described what a tool should do in plain language, mostly in German. I worked in the GitHub Copilot chat in VS Code. According to my chat histories, mostly with Claude (Opus models), occasionally with GPT and Gemini. These histories also cover other work and are not assigned to individual tools.

  2. Letting the AI write the code, with fixed rules

    ClaudeHTML/JavaScriptSheetJSJSZip

    The AI wrote and changed the code. A binding guideline in the repository is explicitly addressed to AI assistants as well. It covers versioning, the change log, the offline principle and a consistent look. In the browser tools, Word templates are filled in directly at XML level, so Word itself is not needed.

  3. Testing with real documents

    WordExcel

    I tested each tool on real documents for subject-matter correctness and refined it step by step. New features are first labelled “test phase”.

  4. Versioning

    GitPull requests

    Git with a protected main branch and pull requests. Every tool has its own change log and shows its version.

  5. Delivering everything in one file

    PowerShellHTML/JavaScript

    A build step in PowerShell packs the portal, the tools and the guides into a single HTML file. Shared libraries are embedded only once and compressed, and the tools run in isolated frames. The file starts with a double-click; no server is needed.

Planning something similar?

Let me help you.

Request an intro call

Next case studyAn animated commercial