Mahdi Ahmadi

Developed Python and HPC pipelines for more than 70,000 wavelength-geometry samples from Lumerical FDTD simulations. Engineered magnitude- and phase-aware representations, applied dimensionality reduction and clustering, and developed low-rank complex-field and shared-pole models for photonic-mode analysis. Used blocked cross-validation, sensitivity analysis, and robustness testing.

At the BTU AI Lab, in cooperation with Lebenshilfe e.V., developed a YOLO-based computer-vision system for real-time detection and counting of screws in assembly trays. Prepared and annotated the image dataset, automated sub-tray extraction, and evaluated detection and counting robustness under occlusion and reflections using reproducible experiments.

Built reproducible Python and PyTorch benchmarks to evaluate quantized large language models under controlled hardware-induced weight perturbations and drift. Modified model checkpoints to simulate systematic parameter drift and evaluated inference behaviour, robustness, reliability, and sensitivity across controlled experiments.

Developed a modular Python toolbox for Raman spectroscopy and diffraction data. Implemented baseline correction, cosmic-spike removal, peak detection, and peak fitting, then built workflows for dimensionality reduction and spatial clustering of Raman maps. Delivered a PyQt interface for visualization and iterative validation with materials-research collaborators.

Master’s thesis submitted August 2026; defence scheduled September 2026. Thesis: Identification and Classification of Photonic Modes in Complex Nanostructures Using Machine Learning.