I build Agentic AI Systems
Data Engineering & AI student at INPT Rabat, obsessed with intelligent agents, LLMs, and turning messy data into meaningful systems.
I'm a Data Engineering & AI student at INPT (Institut National des Postes et Télécommunications) in Rabat, Morocco. My focus lies at the intersection of Agentic AI, LLMs, and distributed systems— building systems that don't just process data, but reason about it.
Before INPT, I completed two years of intensive CPGE MPSI preparatory classes at Lycée Mohamed V in Casablanca, where I ranked 225th out of 3,000+ candidates nationally in the CNC competition and earned admission to CCINP schools in France with a scholarship.
When I'm not training models or architecting pipelines, I'm learning Spanish (the language, but also building AI tutors for it), and exploring how blockchain can make ML systems more robust.
A curated overview of the tools, frameworks, and languages I work with to build production-grade AI and data systems.
A selection of projects that demonstrate my range across AI, data engineering, and distributed systems.
Fine-tuned a 7B-parameter LLM on 2,000+ curated Spanish language samples for conversational tutoring. Architected a RAG pipeline integrating 500+ language resources with TTS/STT modules for fully voice-based practice sessions.
Developed an agentic AI system using the Model Context Protocol (MCP) to automate 5 CRM workflows including lead scoring, client segmentation, and follow-up scheduling. Reduced manual data entry by 60% and improved query response time by 3×.
Orchestrated on-chain majority voting across 50 decision trees via 3 Solidity smart contracts. Stress-tested the system against 4 adversarial attack vectors to validate robustness of the distributed ML approach.
Architected a distributed monitoring platform coordinating 8 concurrent agents across 3 protocols (TCP, UDP, RMI) with automated failover recovery. Built for real-time node health tracking in dynamic network environments.
Built an end-to-end adversarial ML system on the GTSRB dataset (43 classes, 50K images). Implemented FGSM and PGD attacks from scratch in PyTorch, reducing a 97.6% baseline CNN to 19.6% accuracy under attack. Designed a minimax adversarial defence (PGD-in-the-loop) recovering adversarial accuracy to 64.7% while retaining 86.4% clean accuracy. Produced 15+ diagnostics: saliency maps, t-SNE feature space analysis, per-class vulnerability charts, and pixel-level perturbation heatmaps.
Scraped and cleaned 10,000+ records from OpenFoodFacts using Scrapy. Benchmarked 4 ML models and selected XGBoost at 87% accuracy as the optimal classifier for food product categorization.
My academic path and the milestones that shaped my technical perspective.
INPT (Institut National des Postes et Télécommunications) · Rabat, Morocco
Lycée Mohamed V · Casablanca, Morocco
Lycée Bismillah · Settat, Morocco
I'm currently open to internships, research collaborations, and interesting projects. Whether you have a question or just want to say hi — my inbox is always open.