Mohammadali Shakerdargah

• Deployed an information-theoretic method for reducing hallucinations in CoT reasoning models. • Enhanced Mistral-7B accuracy by 4.8% by applying token reframing under high-entropy conditions. • Led strategy discussions with headquarters and global teams on efficient LLM deployment for edge.

• Improved Llama-3 end-to-end latency by 6.1% by developing and integrating MAS-Attention. • Achieved 2.75× speedup and 54% less energy than SOTA methods on T5, ViT, and BERT variants. • Deployed MAS-Attention to 10M+ users by integrating it into Huawei DDK. • Achieved 31.15% faster attention inference via pipelined heterogeneous computation on edge NPUs. • Enhanced throughput on resource-constrained NPUs using stream-based parallel attention kernels.

Achieved 30% latency reduction in Stable Diffusion 1.5 UNet, optimizing each iteration from 0.6s to 0.42s using TVM Unity’s dynamic shape support and hardware-tuned kernels.

• Boosted diagnostic accuracy by 7% with fused medical text and X-ray images using BERT-UNet model. • Achieved 87% accuracy on sentiment analysis of imbalanced COVID-19 tweets using fine-tuned BERT. • Developed scalable multimodal pipelines across diverse healthcare datasets.

• Achieved 93% accuracy on Sign-Language classification using LSTM based hybrid model and MediaPipe. • Collected and curated 6K+ video samples to build Persian Sign Language recognition dataset. • Built real-time translation pipeline optimized for inference latency and gesture precision.

Engineered and operated a precision Microfluidic-Robot for targeted drug delivery, leveraging STM32 microcontroller programming to ensure accuracy and control in biomedical applications.

Designed and deployed a silicone-based Soft-Gripper Robot with pressure-controlled deformation for flexible actuation in Soft Robotics

Developed a simulated industrial control system using PLC programming, automating temperature and pressure controls for an assembly line to optimize efficiency and reduce manual intervention.