Raphaël Benz

– NLP data-scientist/ML-eng.

Work on several projects: – Development of SpreadAuto, a large-scale financial document parsing tool extracting datapoints from balance sheets, income statements, cash-flow statements and annexes (hundreds of pages). Built a full non-LLM pipeline combining YOLO, Camelot, OCR, parsing rules and dictionary matching. Tool deployed in production and used by 30+ daily users across Risk & Finance. – Implementation of TDD, clean architecture principles, GitLab CI/CD, document-level testing framework and monitoring. Continuous improvement of extraction quality through model optimisation and automated validation. – Industrialisation of ML pipelines: Docker packaging, on-prem deployments, secured API interfaces, GPU/CPU resource management and automated release processes. – Development of a RAG-based extraction system, document-type agnostic, with custom modules for termsheets, valuation reports and energetic diagnostics. Integration of internal embeddings and vector search. – Deployment and optimisation of on-prem LLMs under strict RGPD constraints: quantization, inference service construction, secure APIs ensuring that sensitive financial data never leaves BNP systems.

Work on several projects: – Business potential model for La Plateforme Du Bâtiment, predicting future evolution of client visits both for churn and positive development patterns based on BG-NBD model and several other transformed client data. Model on-going production stack, and on-going ROI measurements. Technical performances measurements showed strong signal captation. – Departmental market shares computation with linear models, based on regional market shares + internal and INSEE data. Promising performance measurements based on proxy patterns correlation - as direct measurements were impossible for this unsupervised learning task -. – Segmentation of client base for DSC (Distribution Sanitaire et Chauffage) and construction of feedback-loop labellisation to be used by commercial forces for continuous learning/monitoring/performance enhancements. – Predicting business potentials of client base - embedding, clustering, supervised-learning - to drive commercial actions. – Creation of consumption embedding of client base based on auto-encoder to generate explicative features increasing performance metrics in several productionized ML models (churn/segmentation/business potential ...). – Developpement of testing and monitoring architecture for productionized projects.

– Engineered Python-based data science programs for sound analysis and vehicle classification, deploying algorithms on connected acoustic sensors in various European cities. – Leveraged CNNs applied to melspectrograms and GCC-PHAT to achieve accuracy rates exceeding 90%. – Developed a specific model for bird species inventory based on mel-spectrogram analysis, successfully deployed in several natural parks across France. – Built a comprehensive data science pipeline for sound-based analytics, including vehicle detection, classification, and bird species identification. – Incorporated a robust testing and monitoring system to ensure optimal performance of productionized models.