Writing
Write-ups on design engineering, AI products, 3D tooling and accessibility. Mostly about the load-bearing details: the bug that was not where I looked, the default that quietly did the opposite of what it promised, the unglamorous part that decides whether a thing actually works.
My 3D compressor had been shipping uncompressed geometry since the day it launched. The encoder never loaded, the loader swallowed the 404, and every export still said "done". A story about tools that lie by omission, and the one assertion that would have caught it.
Every image model I ran carried "deformed" and "ugly" in its default negative prompt. For medical imagery that is not a quality setting, it is an instruction to delete the finding. What looked like a censored model was my own config fighting me.
Two days, three confident diagnoses, three fixes that all "worked". They worked because two thirds of the time anything works. What finally broke the loop was not a better theory, it was counting 39 launches already sitting in the logs.
I asked a local model who owns each domain my network talks to. It answered every one with total confidence, including the six domains I invented. The eval that mattered was not the real inputs, it was the ones that do not exist.
Building karaoke read-aloud in Lens - syncing a spoken word to its exact spot on screen, inside a document you can still edit, for readers with dyslexia.
Nothing under that tag yet.