Mourad Yahia

• Refactored and standardized data projects across the group to reduce technical complexity and improve maintainability. • Implemented a data platform based on the Medallion architecture to structure data flows and analytical use cases. • Developed internal MLOps tools for data quality checks, data drift monitoring, artifact versioning, and processing reports. • Designed and deployed CI/CD pipelines including unit tests, integration tests, and automated deployments. • Defined technical standards, authored reference guides, and trained teams on data and MLOps best practices. • Developed a RAG chatbot over internal project documentation and Confluence, and built MCP integrations. • Technical environment: Dataiku, Spark, Great Expectations, Whylogs, Pydantic, MLflow, GitHub CI/CD, Kubernetes, LangChain.

RAG: • Built the ingestion pipeline for internal documents (multi-format parsing, structured chunking, metadata enrichment) and vector indexing in Milvus, with incremental reindexing and collection versioning. • Developed the RAG chain in LangChain: hybrid retrieval (dense Milvus + BM25, RRF fusion), query rewriting, cross-encoder reranking, entitlement-based filtering, and grounded answer generation with source attribution. • Set up continuous pipeline evaluation (Recall@K, MRR, faithfulness via RAGAS) and productionized the service through a containerized API, CI/CD with regression tests, and monitoring of p95 latency and token consumption. Recommendation system: • Assisted the client in defining the needs and designing a solution that centralized product recommendation processes. • Collected data from different sources of the group’s information system (Datalakes, DBs). • Implemented a hybrid product recommendation engine (content-based, collaborative filtering, and time-aware) for customer advisors. • Industrialized the recommendation engine and its training pipelines. • Evaluated the model's performance (3.5 times better than the previous solution) and assessed biases. • Implemented continuous-training pipelines for the model with automatic validation. • Developed an explainability module inspired by the surrogate-model method. Technical environment: PySpark, Python, PyTorch, Kedro, MLflow, GitHub Actions, Teradata, Hive, Linux, Langchain, Milvus, RAGAS, Docker.

• Geolocated magazine articles for a French market leader in the trade press. • Studied the state of the art of named entity recognition (NER) methods and technologies. • Indexed the Geonames gazetteer database on an ElasticSearch node. • Implemented the data access layer and query refinement for toponym resolution. • Retrieved data and conducted Exploratory Data Analysis (EDA). • Studied the state of the art of recommendation models. • Implemented a magazine article recommendation engine based on reading history (reading sessions with a fading coefficient) and article similarity (content-based). Technical environment: Python, Pycharm, Tensorflow, Transformers, Docker, Flask, Databricks, Spacy, Solr, ElasticSearch.

Customer segmentation based on mobility survey data. • Raw data preprocessing, Exploratory Data Analysis and feature engineering. • Selection of active variables. • Application of dimensionality reduction models (ACM, CCA, PLS regression). • Segmentation and stabilization using different clustering models. (CAH, Kmeans,..) • Evaluation of segmentation stability. Technical environment: Python scientific toolkit, Gitlab, Git, Jupyter.

• Assisting the client in setting up a solution for the evaluation of billboards through geographic data and multimodal traffic. • Collection, structuring and construction of different databases (clients/ POI/ trips). • Project structuring (code modularization) and integration of different components into the codebase. • Pre-processing and exploratory data analysis (EDA) of multimodal traffic flows. • Modeling and prediction of the travel vectors for the french population (Speed, Distance, Schedules, modes, etc.). Technical environment: Pycharm, Python, PgSQL, numpy toolkit, gitlab, Git, Jupyter, Docker, tensorflow, shapely

• Supporting the client in setting up a global customer segmentation (Europe+China) based on mobility data from the questionnaire (types of trips and mobility styles). • Exploratory analysis and adaptation of data schemas for machine learning models. • Selection of clustering variables (active) and validation variables (passive). • Application of dimensionality reduction and segmentation models by different clustering models. • Stabilization, validation, and presentation of the segmentation to the marketing teams. Technical environment: Python, numpy toolkit, gitlab, Git, Jupyter.

• Pre-processed and cleaned unstructured data scraped from the web, enhancing data quality and usability. • Automated the extraction of keywords for online commerce competitiveness assessment using Natural Language Processing (NLP) techniques. • Contributed to internal R&D projects, including research and the preparation of scientific publications. • Developed and implemented a chatbot using a sequence-to-sequence (seq2seq) architecture. Keywords: Machine Learning, Embedding, Clustering, TF-IDF, Deep Learning, Word2vec, RNN, TensorFlow, NLP, Python, NLTK, Regular Expressions.

• Conducted in-depth research in Natural Language Processing (NLP), focusing on word embedding techniques to enhance the capture of semantic relationships, including synonyms and antonyms. • Integrated semantic information with contextual data learned by existing neural embedding systems, such as Word2vec and FastText, to improve language models. • Studied and implemented landmark NLP models, including the Neural Probabilistic Language Model (NPLM) by Y. Bengio and the word embedding techniques by T. Mikolov. • Designed and developed neural network models, including a Triplet Network, to explore novel approaches in word embedding. • Implemented evaluation measures and gained proficiency in interpreting and analyzing the statistical results of machine learning models.