At Samsung R&D Institute India, I build production AI systems for smartphones, including on-device MoE LLM inference, battery RUL/SoH forecasting, and audio and image models.
I architected a Llama-MoE inference pipeline for Android and led a three-engineer team targeting 32 tokens per second under mobile memory and compute constraints. I built expert prediction, prefetching, and caching to reduce expert-loading latency, and adapted the pipeline for Qualcomm Hexagon NPU.
I also architected battery Remaining Useful Life and State-of-Health modeling approaches, addressing dataset shift, battery degradation variability, and limited proprietary data. My work combines battery physics with engineered features and degradation-based clustering to improve generalization and reduce deployed model count.
Earlier, I built mobile platform and runtime software across Windows Mobile, Symbian, and Brew at Comviva Technologies. I bring 23 years of software engineering experience, mentor engineers, and use Python, PyTorch, TensorFlow, Hugging Face, Android, AWS S3, Git, and Perforce to take constrained-device systems from architecture through deployment.
