How I use it for open-source pharmacometrics
September 23, 2026
xgxr::StatSummaryBinQuant is a hand-copied fork of ggplot2’s stat-summary-bin.R, pinned in its roxygen block to ggplot2 commit 351eb41. It never picked up a transform step ggplot2 added later.breaks <- breaksbreaks <- scales[[x]]$transform(breaks)Continuous Integration (CI) via GitHub Actions publishes the site.
Allows for workflows, teaching material and tooling can be developed and shared without exposing real clinical data.
This is still a work in progress, with three different synthetic data generators explored so far, each with their own strengths.
The above lets me tell if the syntehtic data looks reasonable.
Once I have an algorithm that works well and the process as acceptable by data privacy, I can validate the code. Thus far, it’s all been “vibe-coded” with Claude Code.
It’s very helpful having a working prototype for building my own understanding for discussing with others.
Does intra-patient dose escalation reach a confirmable active dose sooner in a first-in-human T-cell engager trial?
MCLA-117: a published within-patient ladder, 25 µg to 400 mg, transcribed from the EHA 2020 poster.
A specification, a reading queue, and a references page marking, for every source, whether the claim drawn from it had been checked against the source.
I’m reading/thinking more about the difference between “vibe-coding” vs “agentic-engineering”, the latter meaning to think more about the analysis plan before building a tool. Guide for creating working specifications
pkgdown. For anything else, use a site made by quarto_render().I’ve set up the TrinityMetrics repository for this.
/remote-control.Running Claude directly at claude.ai is blocked until you have done the short internal training that is linked to whenever you hit the block.
claude.ai and do the minimal training required so the site isn’t blocked./remote-control in the chat window to run from other machines, including your phone.[live]