Sushil Dubey — AI Researcher

I’m an AI Software Engineer with 8+ years of experience building AI frameworks, full stack deep learning system and performance‑optimized AI training & Inference. Over the past five years at Intel (Habana Labs), I’ve contributed across the full AI software stack—developing PyTorch device plugins, writing deep learning kernels (GEMM, convolution, attention, MoE), enabling large‑scale distributed training, and improving end‑to‑end model performance on Intel Gaudi hardware. I’ve played a key role in enabling and optimizing open‑source LLaMA, DeepSeek and Stable Diffusion models on Intel Gaudi using 3D parallelism (tensor, pipeline, data, and expert parallelism) with Megatron‑LM. I also led four MLPerf training and inference submissions on Intel AI accelerators, achieving good performance improvement. Before Intel, I worked at Synopsys, contributing to HDL compiler development for FPGA synthesis in C++. Earlier at TIFR (CERN CMS project), I developed GPU‑accelerated parallel clustering and accelerated path‑finding algorithms for high‑throughput CMS detector data processing. I hold Master’s in Artificial Intelligence from the Indian Institute of Science (IISc), with strong mathematical grounding in generative models, LLMs, Diffusion model. I also bring strong problem‑solving skills and solid fundamentals in data structures, algorithms, C++, Python, PyTorch internals, AI frameworks, and experience working with large, production‑grade codebases.

Stackforce AI infers this person is a highly skilled AI Software Engineer with expertise in deep learning and compiler optimization.

Location: Bengaluru, Karnataka, India

Experience: 9 yrs 11 mos