Notes
Notes and essays.
Things I'm thinking about — engineering, ML, and adjacent topics.
How do you alert on something that's supposed to be quiet?
Most monitoring assumes a steady signal, so absence means failure. Event-driven pipelines break that assumption — here's what it took to alert on one without drowning in false positives.
Why our uncertainty measure was wrong
ProofShape's original uncertainty calculation put generative-completion spread in the denominator for observed surface — which quietly declared every observed region infinitely precise.
An evidence-based decision not to fine-tune
When accuracy plateaus, fine-tuning is the default next move. Three independent signals from TruPhoto said not to — and what to check before assuming more training will help.
Accuracy went down four times in a row
What a falling accuracy number across four iterations actually meant for TruPhoto — and why a portfolio full of only-rising numbers should make you suspicious.
How my view on AI-written code changed in six months
In March I wouldn't ship a line I couldn't explain. By September I thought almost all code should be AI-generated — what changed, and what stays the human's job.