Muhammed Abdullah Shaikh

Reduced Samsung eMMC (1.8V rail) sleep current from 52.1 mA to 0.11 mA via low-level driver optimization and power-state tuning; validated robustness with 15 TBW burn-in testing. Hardened ATCU4G FOTA pipeline with CRC-validated headers and fail-safe metadata, ensuring reliable OTA updates over TCP/IP and preventing bricked field units. Upgraded CAN-based ECU Service Tool (UDS bootloader) and recovery workflows, enabling restoration of 400+ field units, saving significant support costs. Designed and implemented a preemptive RTOS scheduler on Renesas RH850 with interrupt-driven context switching and task quantization for real-time determinism. Performed complete board bring-up for RH850 U2A8 MCU (peripherals, clock tree, NVIC/interrupt routing, HAL, and drivers). Automated end-of-line validation of device settings and configurations using Qt-based utility, achieving a ~30% reduction in processing time per device. Streamlined build & release workflows, reducing prep time from 30 to 5 minutes and improving QA throughput. Built reproducible QEMU/Docker environments for debugging and cross-team onboarding, ensuring deterministic firmware tests. Implemented DoIP on LwIP TCP/IP stack for Telematics Control Unit Ethernet variant, enabling secure remote diagnostics. Delivered embedded systems & debugging training to GETs and new hires, covering RTOS primitives, coding guidelines, and Python3.

Explored coreset selection strategies (greedy, k-center) for efficient data sampling in ML pipelines; evaluated performance on SVHN, CIFAR-10, and CIFAR-100 datasets, achieving reduced training overhead with minimal accuracy drop. Analyzed gradient approximation quality across varying sample sizes and training stages, providing insights for scaling deep learning models to large datasets.

Developed a face-recognition system for contactless door unlock, achieving 81.4% accuracy on the LFW dataset using FaceNet and ArcFace with ResNet backbones; contributed dataset class code to PyTorch Vision. Implemented deep learning architectures including Siamese and Prototypical Networks for few-shot classification and U-Net/YOLO pipelines for segmentation and detection on PASCAL VOC 2012 (83.1% / 70.3% accuracy). Built robotic navigation pipelines in ROS: designed PID controllers with Hermite curves, integrated SLAM toolbox with LIDAR sensors, and executed waypoint navigation on Kobuki TurtleBot2. Supervised and mentored 40+ lab members, conducting structured workshops on programming, embedded systems, robotics, and computer vision, while reviewing code and guiding research documentation. Organized and led institute-wide workshops and events, training 200+ students in robotics and AI fundamentals while promoting IvLabs open-source contributions and research visibility. Directed research and project development in AI, robotics, and embedded systems as Vice-Chairman, while contributing directly as a core researcher in computer vision and robotic navigation. Facilitated collaborations with faculty and industry, aligning projects with cutting-edge trends and fostering publication- and product-ready outcomes.