Projects

Aabar

SparkBuild

Pharma Wizard

e-Tayoga

A project about making scholar institutions capable of achieving energetic independence with autonomous and scalable microgrids on production, consumption, and distribution levels using reinforcement learning (RL).

Inspired by the actual state-of-the-art in smart grids, based on much research about renewable energies and university campuses, and optimized with a creative RL reward function, which helps reduce the usage of urban electricity and increase the usage of green energies in a hybrid smart grid with green and fossil sources.

During simulations, our system was able to achieve up to 20% optimization regarding our objective, in comparison to a baseline algorithm that uses a conditional block of code (if/else) with the same purpose.

This project studies Shamir’s Secret Sharing (SSS), a cryptographic method that splits a secret into parts requiring a minimum number to reconstruct it. It explores SSS’s math and security proofs and uses Python to simulate encryption, decryption, and brute-force attacks to test its strength.

The goal is to design a banking authentication system for multi-owner accounts using SSS, ensuring that access requires multiple owners’ approval. This enhances security by preventing single-user access and demonstrates practical applications of cryptography in finance.

A distributed framework for exoplanet survey equipped with machine learning models that could train themselves autonomously using the data they classify based on the innovative self-supervised learning paradigm, and for which we proved the efficiency in our case study.

This project was also presented as a talk at the 4th Annual Meeting of the African Astronomical Society (AfAS) in Marrakech, Morocco.

A short presentation video is available on YouTube.

Inspired by scikit-learn, this package implements several fundamental machine learning algorithms from scratch in Python 3.8, including Single Layer Perceptron, Pocket Perceptron, Adaline, Linear and Logistic Regression, Polynomial Regression, and multi-class classifiers like One-vs-All and One-vs-One. It also includes key learning theory tools such as non-linear transformation, K-fold cross validation, gradient descent, regularization, and theoretical concepts like VC dimension and covering numbers.

Additionally, the package provides utility functions for generating dummy data (2D and 3D) and plotting results, including linear and non-linear decision boundaries and regression lines. All implementations and tools are organized within the lib folder for easy access and use.

A computer vision technique that allows images to be retrieved from a database based on the similarity of their actual visual content, rather than relying on metadata or keywords. This process involves extracting distinctive features—such as color, texture, and shape—from each image and representing them numerically. The system then compares these features using similarity measures to identify and return images from the database that most closely match the query image.

CBIR is widely used in areas like digital asset management, medical imaging, and e-commerce, where efficiently finding visually similar images is essential. By focusing on the intrinsic properties of images, CBIR systems provide a robust and objective way to organize and search large collections based on visual content.

This GitHub repository provides a Python package for unconstrained optimization. It includes implementations of multivariable algorithms like Gradient Descent, Conjugate Gradient, and Newton methods, as well as ten one-dimensional search methods such as Golden Section and Fibonacci. The package also features visualization tools and relies on libraries like NumPy, SciPy, and Matplotlib, along with custom modules for matrix operations, Gauss elimination, and matrix decompositions.

Organized with core modules and utilities, the project offers main scripts to test and compare methods on random functions. This setup creates an interactive learning tool to explore and visualize optimization algorithms effectively in Python.

An innovative approach for detecting overlapping communities in complex networks, such as social networks where nodes can belong to multiple groups. HFGC begins by selecting influential "master" nodes as initial community seeds, assigning them unique labels, while other nodes start unlabeled. Through iterative label propagation, each node updates its community membership vector based on its neighbors, allowing for fuzzy and overlapping community assignments that better reflect real-world network structures.

To improve efficiency, HFGC only updates newly colored nodes and their neighbors in each iteration, making it scalable for large graphs. While the algorithm is notably faster than many existing methods, its modularity scores can be lower, indicating room for refinement. An enhanced version, HFGC+, incorporates multi-criteria decision-making (w-TOPSIS) for better master node selection, though results suggest further research is needed to optimize both speed and community quality.

This Flocking Simulation repository implements the classic Boids algorithm to model and analyze collective bird-like behavior. It provides a framework where individual agents (boids) follow simple rules—separation (avoiding crowding), alignment (matching velocity with neighbors), and cohesion (moving toward the group’s center)—to produce realistic flocking dynamics. The system tracks key metrics such as polarization, angular variance, and kinetic energy to quantify the flock’s coordination and activity.

The repository is organized into modules for defining boids and flocks, running simulations, visualizing behavior in real time, and plotting analytical metrics. Users can customize parameters like flock size and interaction weights, enabling exploration of how these factors influence emergent group behavior.

Time-LLM is a reprogramming framework to repurpose LLMs for general time series forecasting with the backbone language models kept intact. Notably, we show that time series analysis (e.g., forecasting) can be cast as yet another "language task" that can be effectively tackled by an off-the-shelf LLM.

I added a try/except block to launch the model and tokenizer download if not already available locally.

Microsoft Azure is a cloud computing platform provided by Microsoft that offers a wide range of services, including computing, storage, networking, and analytics. It enables individuals, businesses, and governments to build, deploy, and manage applications and services through a global network of data centers. Azure is designed to be flexible and open, supporting various business strategies and stages of digital transformation.

I helped fix a small issue in one of the first tutorials of Azure when using Intelligent OCR services, calling begin_analyze_document_from_url() instead of begin_analyze_document() in the layout model code. 

The EMPIMO Digital Library (GMN) is a collection of digital books created in August 2018 by the Moroccan Team for Preparing International Mathematical Olympiads (EMPIMO). All documents available in this library originate from the EMPIMO Facebook community; I have simply gathered and organized them. This non-profit educational library is freely accessible to everyone under a Creative Commons CC BY-NC-SA license, which allows sharing and adapting the materials for non-commercial purposes with proper attribution and under the same license terms.

The library was developed by me during my high school years with basic HTML/CSS/JS and hosted on GitHub and has had more than 3100 visitors since April 2019.