Roland Baz

• Designed and deployed two production-grade AI applications for the pharmaceutical industry (Atessia’s business teams and the client Novartis), achieving measurable productivity gains of up to 70% • Architecture and development of agentic AI solutions (PoCs), combining multimodal models and advanced prompting with agent orchestration • Technical leadership and project management for the Novartis project, including cross-functional team coordination, sprint organization, code reviews, and alignment with business experts • Worked closely with domain experts to improve application performance by implementing a human-in-the-loop process and creating specialized datasets for application testing and evaluation • Industrialization and cloud deployment via Docker on AWS and Scaleway, in compliance with European data residency requirements, as well as implementation of CI/CD pipelines with GitLab • Implemented a monitoring system for performance tracking (response latency, inference costs, error rates) Tech Stack: LangChain/LangGraph/LangSmith, OpenAI, Anthropic, Mistral, FAISS, AWS (Textract, S3, IAM, Elastic Beanstalk, EventBridge, CloudWatch), PyTorch, HuggingFace, Streamlit, Docker, LLMOps, GitLab

• Design and development of two applied machine learning projects (healthcare and economics), focused on model benchmarking and explainability, in collaboration with researchers and data scientists from Qatar University • Implementation of a complete machine learning pipeline comparing ensemble methods (XGBoost, Random Forest) with classical models (SVM, KNN) to optimize prediction performance • Application of the SHAP framework to interpret models, identify key variables and quantify their impact, ensuring result interpretability adapted to both clinical and economic contexts Tech Stack: XGBoost, Random Forest, SVM, KNN, Regression, Scikit-learn, Pandas, NumPy, SHAP, Imbalanced-learn (SMOTE), Matplotlib/Seaborn

• Developed an AI system for automated scene generation achieving a 50% improvement in coherence compared to the GPT-3.5 baseline model • Architected of an end-to-end LLM pipeline, including fine-tuning of LLaMA 3 (8B) and Mistral (7B) via QLoRA 4-bit/8-bit quantization for efficient GPU training on a theatrical corpus • Implemented of a RAG system with hybrid search (semantic + keyword) to adapt models to the specific style of the playwright • Improved performance via prompt engineering (Chain-of-Thought, few-shot) and validation carried out by business experts. Deployed the application on premises with an A6000 NVIDIA GPU Tech Stack: PyTorch, QLoRA, LLaMA 3, Mistral, LangChain, RAG, Vector Databases, vLLM, Flask, Linux, Git, Streamlit, HTML/CSS/JavaScript

• Collaboration with a team of data scientists and researchers on a research project aimed at analyzing the interaction between financial market assets including cryptocurrencies, bonds, stocks and foreign exchange • Application of Econometric and ML frameworks (GARCH, Quantile Regression) on more than 10,000 daily observations • The research project was published in two top-tier financial journals Tech Stack: Python, Pandas, NumPy, Scikit-learn, Quantile Regression, VAR/GARCH, Matplotlib, Jupyter

• Progressed from Junior to Senior, leading end-to-end delivery of more than 20 projects in the telecommunications sector (cumulative budget > $10M) and managing multidisciplinary teams • Achieved 95% delivery rate and a 20% reduction in project lifecycle through the implementation of Agile/Scrum methodologies • Risk management and executive reporting, ensuring alignment between technical and business teams to guarantee client satisfaction Skills: Project Management, Agile/Scrum, Technical Leadership, Risk Management, Budget Management

Designed and deployed of 4G/LTE architectures and technical quality assurance for more than 10 infrastructures