Muhammad Taimoor

I'm building federated GPU training systems at Skylabs.ai, focused on distributed, cost-efficient large-scale model training. I implement and validate gradient synchronisation and fault-tolerance mechanisms, including resilience testing through node failure and recovery.

Previously at Skylabs.ai, I ran parameter-efficient fine-tuning experiments with LoRA and adapters on HuggingFace Transformers models. I also built evaluation and benchmarking harnesses, retrieval-augmented generation pipelines, and synthetic-data workflows for augmentation, filtering, and automated quality validation.

For my Bio D. Scan final-year project, I trained a custom YOLO11n detector on more than 2,000 self-annotated images and deployed continuous inference on a Raspberry Pi 5 with an AI accelerator. I designed it around real field constraints, balancing detection performance with hardware speed and power budgets.

I've also built deep-learning pipelines for sleep staging and seizure prediction, comparing CNN, CNN-BiLSTM, and transformer models across rigorous data splits. Beyond AI research, I built a React.js, Django, and PostgreSQL transport-booking platform that replaced physical queues for more than 2,300 students and staff.