Sk Abukhoyer

I’m an AI/ML Engineer working at Zentree Labs Private Limited (Client: Qualcomm India Private Limited), focused on making large language models faster, more scalable, and more efficient on real cloud accelerators. I optimize Hugging Face Transformers for Qualcomm Cloud AI accelerators, delivering 20–40% speedups across diverse LLM workloads through custom operator integration and pipeline refinement.

I also expand model capabilities for multimodal GenAI by integrating high-performance pipelines for Diffusion (image/video) and Whisper (audio) within the Efficient Transformers library, improving multimodal throughput while maximizing hardware utilization efficiency. To push scalability further, I architect GPT-OSS Disaggregated Mode support, decoupling compute and memory to improve LLM scalability and resource allocation across multi-accelerator cloud environments.

On the training and reliability side, I build scalable finetuning infrastructure for QAic accelerators with Distributed Data Parallel (DDP) and multi-node training workflows, using memory-optimized PEFT and LoRA techniques. I back this with a Pytest-based unit testing framework and automated CI/CD model validation, reducing integration efforts by 60% and enabling quantized LLM deployment (e.g., Llama-3) for Hexagon Tensor Processors mobile inference with minimal accuracy impact.