loading…
Field notes from production projects — architecture, tradeoffs, lessons. No empty theory.
Quantization-aware training matters more than post-training quant. INT8 RKNN cuts size 3× with <2% accuracy loss. Offline-first saves the day when 4G drops.