Yannis Hadjoudj

Prediction, NLP for Financial Market Agentic Automation for bankers (M&A)

• Fine-tuned LLMs for conversation annotation, reducing processing time by 50% and enhancing accuracy. • Designed a Retrieval Augmented Generation (RAG) framework combined with a LLM for client facing chat-bot with 1M users. • Created a LLM agents for the Quant team that retrieves comprehensive fund information, benchmark funds, and compute comparative indicators such as realised volatility, Sharpe ratio and stress-test VaR • Optimised LLM speed using TGI and continuous batching on a cluster of GPUs. • Engineered prompts to improve LLM accuracy and reduce hallucinations using techniques like Chain of Thought. • Led the red teaming of LLM solutions to detect potential vulnerabilities and hallucinations. • Developed a dashboard to track LLM performance metrics (accuracy, speed, data usage) • Built predictive models for customer lifetime value and churn using advanced methods like Accelerated Failure Time with XGBoost. • Optimized pre-processing pipelines with modern libraries, achieving faster and more efficient data handling. • Conducted R&D on Agentic System use-cases within BNP Paribas ecosystem

• Implemented a supplier risk‐scoring tool in Python using XGBoost on ERP logs and credit‐rating APIs, addressing class imbalance with SMOTE to achieve an AUC of 0.85. • Created a PyTorch Transformer–based demand forecasting model that fused historical orders with a Twitter-scraped “fear index,” cutting RMSE by 20% versus ARIMA. • Developed a real-time Power BI dashboard (DirectQuery to SAP HANA, Azure Data Factory ETL, custom DAX measures) for daily and weekly stock level monitoring. • Orchestrated an Apache Airflow ETL pipeline extracting SAP ECC via OData, transforming data with Pandas/Polars, and loading into a PostgreSQL analytics schema. • Evaluated forecasting models using rolling cross-validation with custom MAPE and weighted MAE metrics, and tracked experiments and drift alerts in MLflow. • Deployed Dockerized FastAPI microservices on Kubernetes and scheduled both batch predictions and model retraining via Airflow DAGs.

->Prediction of production and maintenance delay risks using a supervised machine learning model. -> Development of specific performance metrics to evaluate the effectiveness of the model for predicting delay risks related to maintenance and production. -> Created a data preprocessing pipeline to prepare data for model training and visualization in PowerBI -> Creation of a dashboard for monitoring prediction results in PowerBI. -> Implementation of automation for the continuous collection of data from the internal tool, using techniques such as APIs or web scraping if necessary via Python