Neuroendocrinology and calcium imaging. I build small, open-source instruments for measuring things in neuroscience data — and for finding out when the measurement doesn't hold.
The three instruments run in the browser — no install, no account, and nothing you load is ever uploaded anywhere. The fourth is the review process that keeps the other three honest.
| Colonel Kernel | Recovers the shape that links action potentials to a calcium signal, from ground-truth spike times and a measured dF/F₀ trace — and says so when no single shape fits. | kernel.tonydefazio.com |
| no_peak | CLUSTER pulse detection for hormone time series (LH, GnRH, GH). A faithful port of Veldhuis & Johnson, validated point-by-point against both Igor Pro and the original Fortran. | nopeak.tonydefazio.com |
| bugarach | Finds the moments when many cells fire together — and measures how often you're wrong, on a simulation built from your own recordings. Six coordination detectors lifted out of MATLAB. | bugarach.tonydefazio.com |
| murderboard | An adversarial review harness for document deliverables: eleven reviewer roles that try to tear a draft apart before it ships. Vendored into the three above. | murderboard.tonydefazio.com |
A tool that flatters itself is worse than no tool.
An instrument will always return an answer, and the answer is easiest to trust exactly when nobody has checked what it cost. So each of these reports its own failure modes as loudly as its results — no_peak's front page leads with what CLUSTER misses (roughly 40% of true pulses), bugarach scores its detectors against planted events with known times rather than against another detector's opinion, and Colonel Kernel is built to return "no single kernel fits" rather than a plausible one.
The fourth exists to hold the other three to it.
ORCID 0000-0001-7302-7528 · Bibliography
Provenance differs by project, so one line would be wrong for all four:
- colonel_kernel and bugarach began as my own MATLAB code. Claude (Anthropic's Claude Code) did the port, and does the coding now.
- no_peak is a port of Michael L. Johnson's CLUST5 Fortran. I wrote the Igor Pro port; Claude did Igor → web and Fortran → web.
- murderboard is Claude's code to my design.
Throughout: I set the problems, made the calls — including overriding rules the model had
adopted — and merged nothing I hadn't reviewed. Agent commits carry a Co-Authored-By: Claude trailer.