Machine learning engineer · Paris

Abdelouahed Tahri builds models, then tries to break them.

ML research and engineering across LLMs and agents, time-series forecasting, low-resource NLP (Darija, Arabic, French) and graph learning. I fix the evaluation protocol before the first experiment: out-of-sample tests, leakage checks, honest baselines. MSc Artificial Intelligence at CentraleSupélec.

Open to a 5–6 month research / R&D internship from May 2027
Timestral vs RTE

French electricity demand, a day ahead

one week of the pre-registered backtest, 22–28 June 2026

Hourly French electricity demand from 22 to 28 June 2026, between about 35 and 60 gigawatts, with Timestral's median forecast and 10 to 90 percent band, and RTE's official day-ahead forecast. Both forecasts track the daily cycle; over the whole June to August test, Timestral's average error was 808 MW against 1,216 MW for RTE.
808 MWTimestral error (MAE), Jun–Aug
1,216 MWRTE official forecast error
−34%lower forecast error than the French grid operator's official day-ahead forecast
+25%GDP nowcasting accuracy over baselines at the Central Bank of Morocco
121out-of-sample cohorts in a leakage-safe backtesting engine
1,200+tests in an on-device LLM app

NowBuilding forecasting tools with Timestral and ClickHouse ahead of Mistral's hackathon in Paris on 5 December, and benchmarking self-supervised learning on dynamic graphs at the CVN lab.

i. Focus

What I work on, and how.

  • 01

    LLM applications

    Retrieval-augmented generation, agents with memory and tools, and on-device inference with llama.cpp.

  • 02

    Time-series forecasting

    Foundation models like Timestral, GDP nowcasting with Reservoir Computing, and leakage-safe out-of-sample backtesting.

  • 03

    Low-resource NLP

    Moroccan Darija, Arabic and French: heavy code-switching and no standard spelling.

  • 04

    Graph learning

    Self-supervised learning on dynamic graphs, with evaluation protocols fixed before the experiments.

  • 05

    Optimization

    Constraint programming, genetic algorithms and quantum linear solvers.

ii. Selected work

Things I've built, and measured.

Forecasting · Mistral × ClickHouse

Timestral vs RTE: a foundation model against the grid operator

Mistral's Timestral, zero-shot on a laptop CPU, against France's official day-ahead demand forecast. Pre-registered protocol, leakage-safe contexts queried from ClickHouse, SQL metrics, a live next-day forecast and a Mistral-written operator brief traced in Langfuse.

808 MWTimestral MAE
1,216 MWRTE J-1 MAE
76%of days better
  • Timestral
  • ClickHouse
  • Langfuse
  • Mistral API
LLM · RAG

Zelig, a RAG travel assistant

Hybrid retrieval over a curated knowledge base, persistent memory, and a live news-based safety agent with risk scoring.

  • LangChain
  • ChromaDB
  • FastAPI
  • React
On-device LLM

Macaroon

Offline-first journaling app running a GGUF model on the phone through a custom llama.cpp plugin. Encrypted backups, EN/FR/AR with RTL.

  • llama.cpp
  • Kotlin / C++
  • Capacitor
Optimization

Job-shop scheduler

Exact CP-SAT model combined with a genetic algorithm: optimal on the FT06 and FT10 benchmarks, with interactive Gantt charts.

  • OR-Tools
  • Genetic algorithms
  • Flask
Hackathon

ExoVision AI

Classifies Kepler exoplanet candidates and lets users fine-tune the model on their own labelled data.

  • LightGBM
  • Flask
Quantum

QLP-SOLVER

Quantum linear solvers (HHL, VQLS, adiabatic) for elasticity problems, benchmarked against a classical LU solver.

  • Qiskit
  • FastAPI
  • Docker
Quantum · Club platform

Quantum Chess Arena

Multiplayer chess with superposition and entanglement, an AI opponent and tournament brackets, built for the Centrale Quanta club.

  • FastAPI
  • SQLAlchemy
  • JavaScript
Everything else

More on GitHub

Quantum ML notebooks, coursework and experiments.

github.com/qudotcom

iii. Research

Questions I'm chasing.

CVN lab, CentraleSupélec · 2026–2027 · target NeurIPS / ICLR

Self-supervised learning on dynamic graphs

A reproducible benchmark for contrastive and generative dynamic-graph SSL models, and a synthetic graph generator with tunable temporal heterophily, smoothness, feature noise, sparsity and unseen nodes, to isolate what drives each performance gap.

  • PyTorch Geometric
  • Slurm
  • Node classification
  • Link prediction
Preprint · submitted to the Journal of Asset Management

Systematic strategy backtesting and validation

A reproducible Python backtesting engine with explicit data-leakage protection and out-of-sample evaluation across 121 cohorts. Read it on SSRN.

iv. Experience

Where I've shipped.

  1. May – Aug 2026

    ML Intern, Conversational AI

    CRSC, deployed at BMCI Bank

    Owned the Moroccan Darija language component of a voice AI system that trains bank sales teams.

  2. Jul – Aug 2025

    Data Science & Modelling Intern

    Bank Al-Maghrib, Central Bank of Morocco

    GDP nowcasting with Reservoir Computing: 25% more accurate than classical statistical baselines. Results presented to senior management.

  3. 2025 – present

    President, Centrale Quanta Club

    École Centrale Casablanca

    Leading student research projects, technical conferences and machine learning workshops.

v. Education

Trained in maths first.

2026 – 2027

MSc Artificial Intelligence

CentraleSupélec, Université Paris-Saclay

2024 – 2027

Diplôme d'Ingénieur, AI & Digital Transformation

École Centrale Casablanca

2022 – 2024

CPGE MP

Lycée Omar Ibn Al-Khattab, Meknès

  • Arabic native
  • French C1
  • English professional
  • Pre-Plancks physics competition: 4th (2024), 8th (2025) nationally