About Me

I’m a Scientific Research Engineer working in Natural Language Processing, mostly on model efficiency and multilingual systems. My background mixes research (published work on NMT compression and efficient Arabic language modeling) with full-stack development and systems/infrastructure work — so I can take a model from experiment to production myself instead of handing it off and walking away.

Professional Profile

My research centers on efficient neural machine translation, efficient language model pretraining, and computational morphology, validated on Arabic, a morphologically rich, low-resource-tooling language that stress-tests methods most work only validates on English or French. The techniques generalize by design. I’ve also built and deployed local LLM infrastructure (Ollama-based) with retrieval-augmented generation pipelines in production, and I’m hands-on with the underlying systems: administering Linux servers and GPU workstations, containerizing services with Docker, and managing deployment end-to-end, from model to served application.

Research & Publications

Recent work below, more on my publications page.

  • Efficient Multilingual Neural Machine Translation via Corpus-Driven Vocabulary Pruning: An English-Arabic Case Study (arXiv, 2026): a general optimization framework combining corpus-driven vocabulary pruning with targeted fine-tuning for multilingual NMT models, evaluated on M2M100, NLLB-200, and mBART-50 for English-Arabic. The approach cuts vocabulary size from over 128,000 to roughly 10,000 tokens, yielding a 60% memory saving with no loss in performance, and the pruned/fine-tuned M2M100 model matches or beats a dedicated bilingual baseline on BLEU and COMET. Read the paper
  • AraSSM: A Bidirectional State-Space Encoder for Arabic Masked Language Modeling (arXiv, 2026): introduces AraSSM, a bidirectional Mamba encoder pretrained via masked language modeling on Arabic Wikipedia and CulturaX, trained entirely from scratch on four consumer-grade RTX 2080Ti GPUs over about ten days. As a linear-time alternative to attention-based encoders like AraBERT, MARBERT, and CAMeLBERT, it matches or exceeds published Transformer baselines on Arabic sentiment classification and is competitive on named entity recognition and extractive QA, despite the modest hardware budget. Read the paper
  • A General Framework for Efficient Multilingual Neural Machine Translation: a pruning-based framework for reducing multilingual NMT model size and inference cost while preserving translation quality (validated on English-Arabic).
  • Transformer + VQ-based Arabic Root Extraction (in progress) View all publications

Research & Interests

I keep up with the field the slow way — reading papers, not headlines:

  • Scientific watch: following new work on efficient training, NLP, and LLMs as it comes out.
  • Reviewing: regular reviewer for NLP venues (ACL, EMNLP) and arXiv reading beyond that.

Expertise

  • NLP & ML: Neural Machine Translation (OpenNMT, Fairseq), model pruning/compression, Transformer and state-space (Mamba) architectures, computational morphology, sentiment analysis (BERT, AraBERT), textual similarity (FAISS).
  • LLM Systems: Local/on-prem LLM deployment (Ollama), retrieval-augmented generation (RAG), vector search.
  • Engineering: PyTorch, REST API development (FastAPI, Django), Frontend (React, TypeScript).
  • Systems & Infrastructure: Linux server and GPU workstation administration, Docker containerization and app deployment, CI/CD-style deployment workflows, Database management (PostgreSQL, MongoDB).

Languages

  • French: Native/Fluent
  • English: Professional
  • Arabic: Native/Fluent

“A model that only runs in a notebook isn’t finished work.”