MSc Artificial Intelligence @ CentraleSupélec (Paris) · Engineering @ École Centrale Casablanca
I build and evaluate ML systems end to end: LLM agents and RAG, time-series forecasting, low-resource NLP (Moroccan Darija, Arabic, French) and graph learning. My habit is to fix the evaluation protocol before running experiments: out-of-sample tests, leakage checks, honest baselines.
🎯 Open to a 5–6 month research / R&D internship from May 2027 · 🏁 Hackathons welcome
| 📈 Forecasting | Timestral vs RTE: Mistral's Timestral beats the French grid operator's official day-ahead demand forecast (808 vs 1,216 MW MAE) in a pre-registered backtest on ClickHouse. Also GDP nowcasting at the Central Bank of Morocco (Reservoir Computing RNN, +25% accuracy over statistical baselines) and a leakage-safe backtesting engine evaluated out of sample across 121 cohorts (SSRN preprint) |
| 🤖 LLM apps & agents | Hybrid RAG with persistent memory and a live monitoring agent (Zelig), and on-device LLM inference with llama.cpp on Android (Macaroon) |
| 🗣️ Multilingual NLP | Owned the Moroccan Darija component of a production voice-AI system: heavy code-switching and no standard spelling |
| 🧮 Optimization | CP-SAT combined with genetic algorithms, optimal on the FT06/FT10 job-shop benchmarks (live demo) |
| 🚀 Shipping fast | Hackathon MVP (ExoVision AI), Dockerized APIs, deployed demos, tested code (1,200+ tests in Macaroon) |
Self-supervised learning on dynamic (temporal) graphs, CVN lab, CentraleSupélec (2026–2027)
- A reproducible benchmark for contrastive and generative dynamic-graph SSL models, with the evaluation protocol fixed before any experiment
- A synthetic graph generator with tunable temporal heterophily, smoothness, feature noise, sparsity and unseen nodes, to isolate what drives each performance gap
- Node classification and link prediction with PyTorch Geometric on Slurm · target venue: NeurIPS / ICLR
- ML Intern, Conversational AI: CRSC (deployed at BMCI Bank) · May–Aug 2026 Owned the Moroccan Darija language component of a voice AI system that trains bank sales teams.
- Data Science & Modelling Intern: Bank Al-Maghrib (Central Bank of Morocco) · Jul–Aug 2025 GDP nowcasting with Reservoir Computing: preprocessing, feature engineering, tuning and evaluation. +25% accuracy over classical baselines; dashboard and results presented to senior management.
- President: Centrale Quanta Club, École Centrale Casablanca · 2025–present Leading student research projects, technical conferences and ML workshops.
| Project | What it does | Stack |
|---|---|---|
| timestral-vs-rte | Day-ahead French electricity demand: Mistral Timestral (zero-shot, CPU) vs RTE's official J-1 forecast. Pre-registered protocol, leakage-safe contexts queried from ClickHouse, SQL metrics, live next-day forecast, Mistral operator brief traced in Langfuse | Timestral · ClickHouse/chDB · Langfuse · Mistral API · pandas |
| Zelig: tourism-ai | Hybrid RAG assistant: grounded answers over a curated knowledge base, persistent vector memory, and a live news-monitoring safety agent with risk scoring. REST API, Dockerized | LangChain · HuggingFace · Gemini · ChromaDB · FastAPI · React |
| macaroon-app | Offline-first companion running an on-device LLM (llama.cpp via JNI on Android). AES-GCM encrypted backups, EN/FR/AR with RTL, 1,200+ tests | JavaScript · Capacitor · Kotlin · C++ |
| job-shop-scheduler · demo | Exact CP-SAT model combined with a genetic algorithm: optimal on FT06 and FT10, with interactive Gantt charts | OR-Tools · Flask · Plotly |
| exovision-ai | Hackathon MVP: classifies Kepler exoplanet candidates and fine-tunes on user-labelled data in the app | LightGBM · Flask |
| QLP-SOLVER | Quantum linear solvers (HHL, VQLS, adiabatic) for elasticity problems, benchmarked against a classical LU solver | Qiskit · FastAPI · Docker |
| quantum-chess-arena | Centrale Quanta club platform: multiplayer quantum chess with superposition, entanglement, an AI opponent and tournament brackets | FastAPI · SQLAlchemy · JS |
| quantum-ml-notebooks | Variational quantum classifier vs. a classical neural-network baseline | PennyLane · PyTorch |
- Systematic strategy backtesting and validation framework: a reproducible Python backtesting engine with explicit data-leakage protection and out-of-sample evaluation across 121 cohorts. Preprint on SSRN; submitted to the Journal of Asset Management.
LLMs & NLP: RAG · agents · LangChain · HuggingFace Transformers · ChromaDB · llama.cpp / GGUF · low-resource & multilingual NLP Forecasting & ML: time-series forecasting · nowcasting · backtesting · out-of-sample validation · PyTorch · PyTorch Geometric · TensorFlow · scikit-learn · LightGBM · pandas · NumPy Optimization: Google OR-Tools (CP-SAT) · genetic algorithms · Qiskit · PennyLane Engineering: Python · FastAPI · Docker · REST APIs · Slurm · Git · JavaScript
- MSc Artificial Intelligence: CentraleSupélec, Université Paris-Saclay · 2026–2027
- Diplôme d'Ingénieur, AI & Digital Transformation: École Centrale Casablanca · 2024–2027
- CPGE MP: Lycée Omar Ibn Al-Khattab, Meknès · 2022–2024
🌍 Arabic (native) · French (C1) · English (professional) · 🏆 Pre-Plancks physics competition: 4th (2024) and 8th (2025) nationally
