Built distributed pretraining + fine-tuning pipeline using DeepSpeed and vLLM. Optimized inference latency and scaling across clusters.
Designed a novel zeroth-order optimization algorithm with adaptive momentum. Achieved sublinear convergence rate O(logT / sqrt(T)).
Deployed CNN + MFCC on ARM Cortex M4 for real-time speech control.
LSTM-based motion gesture recognition on embedded system (93% accuracy).