I'm an AI and software engineer from Morocco. I work mostly on machine learning and applied maths, and I spend a fair amount of time getting models to run on the hardware I actually have rather than the hardware I'd like.

The part I enjoy most is the checking: does this metric measure what it claims to, and does the result hold up when I change the preprocessing?

Machine Learning · Reinforcement Learning · Computer Vision · Natural Language Processing · Applied Mathematics · Efficient Software

• Adversarial Testing of MedGemma, stress-tests a medical model with adversarial prompt prefixes written by a second model, while a third scores the answers. All three fit in one GPU session thanks to 4-bit quantisation.
• Topic Modeling: LDA vs NMF, runs both models on the same corpus and compares them. The topic count comes from a coherence sweep, and I tried five preprocessing variants first to see which words survive all of them.
• NLP for Job Recommendation, matches a CV against roughly 210,000 job postings using keyword, topic and embedding scores. Every result comes with the reasons behind it and the gaps it found.
• Bandit-Optimised Diabetes Classifier, an MLP whose decision threshold is picked by a multi-armed bandit. Accuracy barely moves, but recall gains about 4 points, which is the trade you want for screening.
• Image Denoising, searches over classical filters and keeps whichever one best restores a noisy image, measured with PSNR and SSIM against the clean original. Bilateral wins on Gaussian noise, median on salt-and-pepper.
• Distributed Smart City Platform, a Java system where the same services are reachable over REST, RMI, raw sockets and WebSocket, with an embedded database behind them and a live dashboard on top.

A Modular DevSecOps Framework for Vulnerability Detection, Risk Scoring, and LLM-Augmented Remediation in CI/CD Environments S. Sadki, A. Lamgari, E. Y. Tebbaa. CloudTech'25, Lecture Notes in Networks and Systems 1941, pp. 213–227, Springer, 2026. doi:10.1007/978-3-032-23844-3_23

Python · C · C++ · C# · Java · SQL · PyTorch · scikit-learn · FastAPI