Notes from the founder on building modern tooling for clinical and tabular data — SAS, DuckDB, WebAssembly, and the craft of byte-parity engineering.
v0.7 of our open-source DuckDB statistics extension adds lm_fit: an OLS aggregate that fits one regression per GROUP BY group, with heteroskedasticity- and cluster-robust standard errors (HC0–HC3, CR0/CR1). Plus a portable bootstrap and a new linear-algebra kernel underneath.
Two RNG bugs, one debugging session: DuckDB broadcasting a single random draw across 2,048 rows, and a standard-library uniform distribution with a clipped tail. Both produced plausible-looking numbers, both are invisible in code review — and both fell to a Monte Carlo checklist, run by Claude Code.
A tour of the-stats-duck v0.6.0: a one-shot data profiler, regression and bootstrap CIs in plain SQL, ten new distributions, and SAS/XPT imports that got 52x faster — with live Bedevere demos you can poke at.
A friendly tour of the SAS Transport (XPT) format: everything you never wanted to know about it, and how to decode its IBM hexadecimal floats.
Notes from the founder on building modern tooling for clinical and tabular data — what we ship, how we build it, and the trade-offs that show up along the way.
