Deep dives into on-device AI, iOS engineering, and the infrastructure behind voice-first personal computing.
Benchmarks narrowed 113 models to a shortlist, but they never picked a winner. Four steps, each born from the last one's failure, that gave every Audria workflow the model that actually fits it.
How Audria detects when a conversation shifts topics in real time — the math behind semantic drift detection, embedding-based segmentation, and why traditional approaches fail for voice-first AI.
We built a complete voice-first AI system that runs its entire pipeline locally on an iPhone — speech recognition, language model inference on the Neural Engine, constrained decoding, and a knowledge graph memory system. All in airplane mode.