Writing
Notes on shipping LLM features and measuring whether they work. RSS
Evaluating an agent that applies for jobs on my behalf
How I decided whether to trust an autonomous pipeline with my own applications — graders, traps, and what "good enough" meant.
Building Armenian NLP when the data isn't there
Notes from building an NLP toolkit for a low-resource language, what worked, what didn't, and how to report results honestly.