Generative AI · Long read

Claude + BMADBuilding critical software with one engineer and an AI

What can generative AI make possible where failure could cause serious harm? Twelve weeks of building a product that includes a medical device, following the process for the highest software safety class, with Claude and the BMAD method: what AI made possible, which practices made use of it, what stayed with people, and how the mistakes made along the way were caught and corrected.

Summary

This article follows one real project from the inside: a clinical software product for hospitals, part of which is a medical device, developed as a whole following the process IEC 62304 sets for the highest software safety class, with Claude as the developer and BMAD as the method. On this project the team was one person, working part-time for twelve weeks: about a third of full time, by the project's records.

This is a report from a gate, on work in progress. What exists after those twelve weeks is a design-control record in progress: requirements traced to tests, a risk register, decision records, verification logs on identified builds, and an open item for every known gap. What does not exist yet matters as much. Parts of the software development file are still templates. One person wrote, reviewed and approved the work; the AI reviewers raised findings, but they are not independent reviewers in the sense a quality system means. No regulatory affairs specialist has signed anything, no clinician has used the product, no notified body has ruled on where the device boundary runs, and nothing has been released.

One line runs through the article: AI opens possibilities, good practices turn them into results, and the team stays at the helm. A volume of work that once needed a team comes within reach of one person; architecture options can be built and tested instead of argued over; the evidence a regulated product needs can be produced with the code. None of them comes guaranteed. Claude writes code fluently and decides badly when left to decide. BMAD adds roles, a chain from product definition to tested story, and a written artefact at every step. A quality management system adds the controls, and the article tells how its wiring failed silently for ten weeks, and what caught it.

The practices are the ones behind every article on this site: Scrum and its walking skeleton, boundaries drawn with Domain-Driven Design, and a business case built gate by gate, each gate a question answered before the next stage is funded. Four have been passed so far: what generative AI is really worth, whether it was worth building software with, which method to put around it, and whether a configurable medical device could work from end to end. When building becomes cheap, some long-held trade-offs reverse, and the people a project needs are no longer the same. The article ends on that change, what it means for development teams, and leaves the rest to a separate article.

This article is written for CTOs and CIOs weighing AI-assisted development for systems where failure matters, for quality and regulatory leaders who will be asked to sign off on it, and for executives and team leads who will have to manage what it changes for their teams.

Short on time? Read the prologue, chapter 01 and the takeaways.