Data Scientist / AI Engineer

• Email: boulaich.mohamed970@gmail.com
• LinkedIn: Mohamed Boulaich
• GitHub: BlcMed

Summary

I’m Mohamed Boulaich, a Data Science & Operations Research engineer with a degree from INSEA. I’m passionate about transforming complex data into smart, actionable insights and designing solutions that actually work in the real world.

I’ve worked on projects ranging from OCR-based document processing pipelines to RAG-powered chatbots, applying software engineering best practices like CI/CD, testing, and clean code. My interests lie at the intersection of machine learning, optimization, and automation — anything that makes systems more intelligent and efficient.

Professional Experiences

• Machine Learning Freelancer - OCR for tables
  Through Independent Consultant, Remote Jul. 2024 – Oct. 2024
  • Developed textify-docs, a Python library for text extraction from diverse document formats, focusing on modularity and reusability.
  • Built a pipeline using OpenCV for image preprocessing, followed by Pytesseract for efficient text extraction.
  • Applied Table Transformer (TATR), based on DETR, to enhance table extraction accuracy.
  • Utilized LLMs to identify and structure relevant information.

The main pipeline for extracting information from the “portail marocain des marchés publics”

The core algorithm behind the textify_docs library

• AI Research Intern (PFE) – Multi-Agent Reinforcement Learning Ai Movement, UM6P, Rabat, Morocco Feb. 2025 – Aug. 2025
  • Designed and implemented a novel QMIX variant leveraging agent contribution masking and regularization, resulting in a compressed model with lower computational overhead for cooperative multi-agent tasks.
  • Conducted a state-of-the-art review of MARL algorithms in shared reward settings and benchmarked against standard environments using PettingZoo and PyMarl.

The QMIX architecture. Each agent has a local network. A central mixing network, whose weights are generated by state-conditioned hypernetworks (red), combines the individual utilities into total Q value while ensuring monotonicity. Original Paper

Architectural comparison. QMIX-Masked introduces a masking layer that selectively prunes agent q-values before it reaches the central mixing network of the standard QMIX architecture.

• AI Engineering Intern - RAG Maroc Telecom, Rabat, Morocco Jun. 2024 – Aug. 2024
  • Leveraged LlamaIndex framework with vector indexing for efficient information retrieval.
  • Integrated ChromaDB for optimized storage and loading of embeddings.
  • Optimized chatbot responses through prompt engineering, hyperparameter tuning, and tools abstraction.
  • Created a user-friendly interface using Streamlit to showcase the chatbot’s functionality.

Streamlit app featuring an interactive chatbot

• Data Analyst Intern Higher Planning Commission (HCP), Tangier, Morocco
  Jun. 2023 – Jul. 2022
  • Cleaned and preprocessed data, handling missing values and normalizing it, then used linear regression to forecast the 2024 population based on the cleaned demographic dataset.
  • Utilized Folium library to create interactive maps for visualizing population distributions across different regions.

Projects

Artistic Neural Style Transfer – PyTorch, Jupyter notebooks, Streamlit (Oct. 2024 - Present)
GitHub Repo

• Utilized Neural Style Transfer with VGG19 CNN for feature extraction and gram matrix computations.
• Developed a Streamlit app allowing users to generate stylized images.
• Fine-tuned hyperparameters to optimize style transfer quality.

Diagram illustrating the Neural Style Transfer process. The content image (bottom left) and style image (top left) are passed through a convolutional neural network. Content and style representations are extracted at different layers. These representations are then used to guide the transformation of a generated image (typically a white noise image), resulting in style reconstructions (top) and content reconstructions (bottom) at various levels of the network. Original Paper

Accent Detection DL model – Pytorch, Pandas, Scikit-Learn, Librosa (Apr. 2024 - Present)
GitHub Repo

• Reviewed literature on accent recognition methods and deep learning architectures.
• Collected audio recordings from the Speech Accent Archive.
• Preprocessed the dataset using silence trimming and noise reduction.
• Implemented and fine-tuned an ANN to predict accents using MFCCs, achieving 0.84 accuracy, 0.79 F1 score, and 0.80 precision.

Sentiment Analysis On Movie Reviews – Spacy, PySpark, Pandas, Scikit-Learn (May. 2024 - Jun. 2024)
GitHub Repo

• Preprocessed text data using tokenization, lemmatization, and vectorization (BoW, Tf-IDF).
• Trained classifiers (Logistic Regression, SVM, Naive Bayes), achieving 86% accuracy with SVM.
• Used PySpark to process large-scale data, improving computational efficiency.

Financial Cointegration Analysis | Time Series Analysis, Python, Jupyter Notebooks
Dec. 2023 – Dec. 2023
GitHub Repo

• Applied the Engle-Granger two-step method and cointegration concept to distinguish between spurious correlations and genuine long-term relationships.
• Analyzed historical price data from Yahoo Finance using Python’s yfinance library.

Testing cointegration between two financial time series

Student Accommodation Clustering | Python, Scikit-Learn, Folium
Nov. 2022 – Dec. 2022
GitHub Repo

• Developed a student accommodation clustering system to suggest optimal housing options based on proximity to preferred locations.
• Conducted data collection and cleaning tasks, including API integration and handling missing values.
• Implemented machine learning techniques, particularly the K-Means clustering algorithm from scikit-learn, to group accommodation options based on their similarity.
• Performed exploratory data analysis (EDA) with visualizations and maps to gain insights into student preferences and housing patterns.

Optimal housing based on students preferred locations.

Education

• Engineering, Data Science
  The National Institute of Statistics and Applied Economics, Rabat, Morocco
  2022 - 2025

• SPE, MP*
  Higher School Preparatory Classes (CPGE) Moulay Idriss, Fes, Morocco
  2021 - 2022

Skills

• Languages: English (TOEFL 82), French (fluent), Arabic (native)
• Statistics: Statistical Inference, Descriptive Statistics, Statistical learning, Machine Learning Methodology, Traditional Modeling, Generalized Linear Models, Time Series Analysis, Hidden Markov Model, Stochastic Processes, Queuing Theory
• Operations Research: Linear and Integer Programming, Stochastic Optimization, Decomposition (Dantzig-Wolfe, Benders, Column Generation), Metaheuristics (SA, GA, TS), Graph Theory
• Libraries: Numpy, Pandas, Scikit-Learn, Pytorch, Transformers, SpaCy, Seaborn
• DevOps/MLOps Tools: Code Versioning, CI/CD (Github Actions), Automated testing (tox, pytest), Docker), Code Quality Tools, Agile/Scrum
• Data Tools: Apache Airflow, Apache Superset