Claude Code at Work
A talk to the group on using Claude Code for open-source work
An informal 15-minute talk at the AI symposium, which meets a couple of times a week for people to show how they use these tools. On how Claude Code is being used for open-source pharmacometrics work. Everything shown is public and none of it touches clinical data.
The format sets the shape: it is a show-and-tell, not a position paper, so the deck carries worked examples and no closing section of caveats.
Documents
- Demo page. The theophylline dataset plotted, a one-compartment model fitted with
nlmixr2using FOCEi, and the fit compared against the observed data. Built as the live-demo target: it renders in about 8 seconds and publishes in about a minute. - Presentation. The slide deck, in revealjs. Four worked examples, why Claude Code rather than Positron, how it is set up, and the parts that need care.
Where the four examples come from
| Slide | Repository | What to point at |
|---|---|---|
| xgxr to CRAN | Novartis/xgxr |
issue #72, the expired-TLS-cert fix, cran-comments.md, version 1.1.6 |
| xgx onto CI | Novartis/xgx |
build-site.yml, the data-checking page |
| synpmx | iamstein/synpmx |
the package and its pkgdown site |
| Intra-patient escalation | this repository | projects/tce-ipde/ specification and references |
Before the talk
rxode2 could not compile a model on this machine until 2026-09-02. R’s FLIBS pointed at a Fortran runtime in /opt/gfortran, which was not installed, so every model build failed at the link step. rxode2 reported it as a missing C compiler, which it was not. ~/.R/Makevars now points FLIBS at the libgfortran and libquadmath that R ships inside its own framework. Installing the GNU Fortran compiler from mac.r-project.org/tools is the proper fix, after which that file can be deleted.
Worth checking the demo still renders on the machine you present from.
What still needs Andy
- The demo itself
- Asking Alex for the training link, noted on the Getting access slide
Cut on 2026-09-02
Three slides suggested in the first draft were removed: What I check, every time, The data boundary, and Where it falls short. The deck now makes its case on the four worked examples and the setup, and says only on the opening slide that none of the work touches clinical data. If you want to try it stayed, as the closing entry path.