00Ident confirmed
01What drives me
I like the unglamorous half of AI work: the ingestion, the queues, the retries. Most of what I build ends in a model call, but the part that decides whether it works is everything before it: backfilling 280k procurement records through a rate limited job queue, checkpointing extracted document text so a failed embedding run resumes instead of starting over, grouping raw solar telemetry into events an operator can actually act on. If it cannot survive being run unattended at 3am, it is a demo, not a system.
The work I am proudest of is a number I made worse. A classifier of mine hit 0.91 ROC-AUC, which felt great until I traced it to study overlap across splits and test set exposure during feature selection. Rebuilt with cohort grouped cross validation and fold specific selection, it reported 0.64. That is the honest number, and finding it taught me more than the first one would have. I would rather ship something defensible than something impressive.
The Mandalorian thing is about craft, not cosplay. A covert of people who are very good at one narrow trade, who maintain their own gear, who take a creed seriously enough to be inconvenienced by it. Beskar gets reforged, never thrown out. That is a fair description of how I feel about a codebase you intend to keep.
Creed
*Keyboard, terminal, debugger.
02Off-duty
Pickup games of whichever one I can find. The best way I know to get away from the keyboard.
Slow, precise, and completely unforgiving of a rushed release. The closest thing I have to meditation.
Books, mostly on paper. A good one will cost me a night of sleep.
Side projects for the fun of it, plus a steady habit of reading papers and trying new models to see what actually holds up.
03Featured work
An MCP tool-call proxy that allows, holds for human review, or blocks each tool call by checking it against what the user originally asked for.
87%recall at a 12% false-positive rate, vs. 50% for a signature baseline
Can gut microbiome profiles tell Crohn's disease from ulcerative colitis across independent cohorts? A reported 0.91 ROC-AUC turned out to be evaluation leakage; the honest number is 0.64.
0.91 → 0.64ROC-AUC before and after removing cohort leakage from the evaluation
04Latest transmissions
Why this site exists, what I plan to write about, and how the whole thing is put together.
Mandalorians forge armour from one unglamorous metal for a reason. There's a lesson in there about picking your stack.
Open channel
Have a job, a project, or a Star Wars hot take?
