"Languages are the best mirror of the human mind, and a precise analysis of the meaning of words would disclose the operations of the understanding better than anything else."
— Leibniz, New Essays on Human Understanding
Arabic-French Translation
Specialized translation, terminology choices and bilingual editorial quality control for scholarly texts.
LLM/VLM Adaptation
Fine-tuning and evaluation workflows for historical handwritten manuscripts and mathematical notation.
Model Output Analysis
Prompt design, response review and error analysis across reasoning, structure and transcription quality.
Python ML Engineering
Data processing, evaluation scripts, reproducible experiments and practical web interfaces.
Research & Development: Multilingual NLP, Document AI & Evaluation
Applied modeling for historical manuscripts, multilingual linguistic data, model output analysis and practical AI systems.
Python R&D
HTR for Historical Maghrebi Manuscripts (Qwen 9B)
Supervised fine-tuning (LoRA/QLoRA via **Unsloth**) of Qwen 3.5 9B on a historical cursive manuscript dataset. Target CER: 13%. (Private model, planned for open-source release).
Maghrebi Dataset Distribution
Total Volume~11K annotated lines
Data Splits
Train: 9,524 lines (85%)
Val: 560 lines (5%)
Test: 1,121 lines (10%)
Source Distribution
RASAM I & II: 8,444 lines
BULAC (Suppl.): 962 lines
Custom Annotations: 931 lines
BnF (Suppl.): 581 lines
MMSH Archives: 287 lines
Modeling & R&D Axes
Masked Autoencoders (MAE): ViT encoder initialization via masked pixel reconstruction (He et al., 2021) to stabilize vision modeling against paper degradation.
Crowdsourcing: Continuous ingestion of verified annotations from Khatt-GPT.
Structure-Aware Recognition of Historical Mathematical Expressions
Historical HMER on Leibniz manuscripts using CoMER, structural supervision and QLoRA-adapted vision-language models.
Research on historical handwritten mathematical expression recognition using the Leibniz-HME corpus. The study compares a compact CoMER recognizer with a 4-bit QLoRA adaptation of Uni-MuMER-Qwen3.5-2B under the fixed Spatial-Constraint evaluation protocol. Manuscript in progress, not yet published.
73.05%ExpRate - adapted Qwen
0.95Mean token edit distance
961Held-out test expressions
43.27%Hard-expression accuracy
Full context image (original manuscript)
Corresponding cropped mathematical expression
Leibniz HMER Adaptation & Evaluation Protocol
Corpus: Leibniz-HME contains 4,571 annotated mathematical expressions from 85 manuscript pages with a normalized 140-token vocabulary.
Protocol: Results use the fixed Spatial-Constraint split: 2,642 training, 968 validation and 961 test expressions, using pre-localized expression crops.
CoMER: Adapted a compact recognizer with external HMER pretraining and auxiliary Symbol Counting and Tree-CoT supervision.
Qwen adaptation: Adapted the public Uni-MuMER-Qwen3.5-2B checkpoint with 4-bit QLoRA and PLAIN, Symbol Counting, Tree-CoT and Error-Driven Learning supervision.
ExpRate is exact expression recognition rate on the fixed S-C test split. Reference-aware oracle analysis is not presented here as deployable model performance.
Scientific Contributions & Conclusions
Historical adaptation: Compared compact CoMER recognition with QLoRA adaptation of Uni-MuMER-Qwen3.5-2B on historical notation.
Structural supervision: Used Symbol Counting and Tree-CoT objectives to expose omissions, grouping errors, decorators and rare historical symbols.
Error analysis: Analyzed residual failures across structural scope, missing symbols and notation-specific ambiguities.
EDL caution: Conditional EDL repaired 27 of 259 rejected PLAIN predictions exactly, but was not reliable as an unconditional post-processing system.
Alignment Visualizer (Leibniz HMER)
y \gleich \frac{x}{x + 1}Example LaTeX overlay - qualitative visualization
Slide to align the original manuscript formula with its LaTeX transcription
Current project focused on scanned Arabic manuscripts: preprocessing, OCR experiments, translation alignment and dataset tooling.
OCRArabic manuscript scans
TranslationArabic-French alignment
DatasetsCuration workflows
Building image preprocessing and OCR evaluation workflows for Arabic manuscript material.
Connecting OCR outputs to translation review and bilingual dataset creation.
Project in active development; no public demo link yet.
Next.js 16+ | Crowdsourcing
Khatt-GPT: Collaborative HTR Platform
Web crowdsourcing platform for collaborative transcription of historical Maghrebi manuscripts. Streamlines pedagogical annotation ingestion to enrich custom datasets.
Interactive application for teaching classical Arabic grammar (I'rab) and translation (SDAIA Saudi Hackathon, Riyadh). Fine-tuned Saudi ALLAM LLM (IBM Watsonx.ai) on proprietary datasets.
ALLAM LLMIBM Watsonx.aiArabic NLPI'rab DatasetSDAIA Saudi
Hackathon Llama 3 | lablab.ai
Arabic Companion: Adaptive Language Learning
Personalized conversational agent for learning classical Arabic. Features real-time grammar feedback, correction loops, and adaptive vocabulary learning. Developed at the Llama 3 Hackathon.
Multilingual NLP engineer working across French, English and Arabic. Builds evaluation workflows, OCR/HTR pipelines, datasets and web tools, with a research background in NLP and document AI.
Education
Université de Nantes, France
M.Sc. in Computer Science — NLP & Smart Computing (ATAL)
NLP, deep learning, Transformers, speech processing, computer vision and reinforcement learning.
Université du Mans, France
Double B.Sc. Major in Mathematics & Economics
Mathematics, statistics, optimization, econometrics and quantitative modeling.
Professional Experience in AI
May 2025 - Aug 2025Scale AI - Remote
AI Training and Model Evaluation Contributor
Reviewed prompts and model outputs for AI training/evaluation.
Checked reasoning, instruction-following, response structure and output quality.
Applied QA rubrics in iterative review workflows.
Independent Projects
Freelance Web Developer & RAG Chatbot Integrator
Built websites, internal tools and AI assistants with React/Next.js, Python/FastAPI and API integrations.
Integrated RAG chatbots for document search, retrieval workflows and domain-specific Q&A.
Delivered end-to-end projects from client need to deployment, content structure and maintenance.
Linux AdministrationVPS (Coolify)Local InferenceData pipelines
IV. Languages & Publishing
French (Native)English (C1)Literary Arabic (Advanced)Arabic-French translation
Translation & Dataset Work
Éditions Héritage Mohammadien (Since 2019)
Translate, write and publish Arabic-French books through an independent publishing house. Handles typesetting, cover design, web publishing and AI-assisted OCR-to-translation dataset work.
Languages
French: Native
English: Professional (C1)
Classical Arabic: Scholarly literacy and oral comprehension
Interests
Chess & Game TheoryLogic PuzzlesTypography & Medieval BookbindingPhilosophy of SciencePedagogical Transmission
