PINHSUAN CHEN

I've built C/C++ software for MediaTek, where I designed a shared-library API that abstracts the Linux TV audio pipeline and enables third-party AI models to integrate with MediaTek SoCs.

I managed real-time PCM audio buffer I/O for voice separation, wrote API specifications and integration guides, and restructured AOSP TV Makefile dependencies to prevent ABI mismatches and silent runtime failures.

My research spans neighbor-guided LLM generation for Bayesian optimization in PyTorch and semantic communication systems. For my dissertation, I built a CUDA-accelerated federated learning pipeline across five edge clients that achieved 99% of centralized-baseline performance on image recovery and classification tasks.

I've also evaluated AI speech codecs, built embedded face-tracking systems on the Maixduino K210, and compressed LoRA adapters to reduce trainable parameters by 98% for memory-efficient deployment.