Marouane Hajri

Contributed to the reduction of detection latency for scalable Neutral Atom quantum computers using an AMD Alveo U55C card. • Trained a CNN for signal detection at a low signal-to-noise ratio. → Reduced the CNN parameter count by 95 % while preserving accuracy. • Developed the CNN core in VHDL and implemented it on the AMD Alveo U55C with host control via Xilinx Runtime (XRT). • Achieved a latency of 2.15 µs per prediction at 400 MHz. → ~4.6× faster at a 20 % lower clock frequency.

- Developed Python/Matlab scripts to automate test procedures and measurements using vector network analyzers and other laboratory equipment - Expanded a long-term drift measurement test bench for vector network analyzers - Measured and analyzed settling times of RF components - Implemented a small FPGA-based controller on Xilinx Zynq-7010 for fan-speed synchronization - Designed, routed, assembled, and tested PCB prototypes - Created interactive Jupyter notebook reports to support technical decision-making

Developed image-processing and deep-learning workflows for light-sheet microscopy scans of mouse peripheral nervous-system. • Designed and trained a guided latent diffusion model to generate synthetic light-sheet microscopy scans conditioned on unlabeled scans and a mathematically generated neuron graph. → Reduced creation of a labeled dataset from 136 h to 17 h (~88 % faster) on 2 A100 GPUs. • Trained on multi-GPU A100 nodes via Slurm (~168 GPU-hours per full pipeline) using PyTorch.

Zweitstudium