AI/ML & Aspiring Systems Researcher | Dynamical Systems · Time Series · Low-Level Systems | Seeking a 2027 Research Internship
I am a final-year Computer Engineering student (M.Eng.) at ENSA Marrakech, Cadi Ayyad University. My work sits between two worlds: low-level systems engineering (C/C++, memory, compilers) and machine learning for signals and dynamical systems. I study how learning systems recover the laws behind a signal, and when they fail.
I enjoy digging into how software works under the hood, from web backends to compiler internals, and I like measuring things carefully before claiming anything about them.
All papers and posters are open access on Zenodo (CC BY 4.0), each with code, data and figures on GitHub.
Can data-driven methods recover the equations of chaotic systems from short, sparse, noisy records? Dense least squares, SINDy, weak-form SINDy, a black-box neural ODE and a sparse-library neural ODE are compared on Lorenz-63 and a four-species Lotka–Volterra model. At about 4 samples per oscillation and 1% noise, the sparse neural ODE recovers the Lorenz equations with a median coefficient error of 0.05, where all derivative-based and weak-form estimators fail.
- Focus: System Identification, Neural ODEs, Multiple Shooting, Uncertainty Quantification.
From Simulated Falls to Freezing of Gait: A Controlled Study of Domain Shift in Wearable Accelerometry
A fall detector trained on healthy volunteers (SisFall) is applied, with no patient labels, to freezing of gait in Parkinson's patients (Daphnet). A CNN-1D, LSTM, GRU and patch Transformer are compared under identical protocols: in-domain AUROC of 0.988–0.999 collapses to 0.49–0.63 under zero-shot transfer, and a training-free freeze index beats every network.
- Focus: Domain Shift, Time-Series Deep Learning, Wearable Sensing, Evaluation Methodology.
A controlled empirical study of five 1D and five 2D algorithms in C++: 740,091 correctness checks with zero failures, exact operation counts in 140 of 140 cases, and fitted exponents of 2.95, 2.01, 1.08 and 1.000. Kadane's algorithm is about 300× faster than the quadratic method at n = 2¹⁰, and in 2D the choice of compression side changes the running time by up to 2,741×.
- Focus: Asymptotic Analysis, Benchmarking, Measurement Design, C++.
- Dynamical Systems & Time Series: System identification, neural ODEs, sparse regression, forecasting under noise.
- Robustness & Generalization: Domain shift, transfer to unseen populations, honest evaluation protocols.
- Systems Programming: Low-level development, C/C++, memory management, performance.
- Algorithmic Theory: Optimization, complexity analysis, empirical algorithmics.
- Compiler Basics: Exploring how code is analyzed and transformed (GCC/GIMPLE).

