Marius Ayrault

‣ Built a production multimodal ingestion pipeline on Databricks: OCR, image-to-text, semantic chunking and embedding at scale. ‣ Architected a two-level multi-agent system with LangGraph: Plan-and-Execute orchestration of ReAct sub-agents, human-in-the-loop checkpoints, SSE streaming via FastAPI and prompt caching on AWS Bedrock. ‣ Designed and ran a RAG evaluation framework (Ragas + LangSmith) across 5 pipeline candidates: the in-house pipeline outperformed the commercial baseline on Precision@k (0.82 vs 0.68) and Recall@k (0.70 vs 0.56). ‣ Worked directly with the founders, from product vision to production: observability and feedback loops in place for data-driven iteration.

‣ Systems delivered from exploration to production: industrialized credit scoring (MLOps), real-time streaming pipeline on Kafka, applied deep learning (vision, NLP, time series), RAG solutions and open-source model deployment. Each project is documented at mariusayrault.com. ‣ An R&D program run independently, structured by a Data Analytics degree at Université Paris 1 Panthéon-Sorbonne.

‣ Co-managed the outsourced production of Rafale maintenance documentation for Dassault Aviation across three teams in France and abroad: technical writing standards, XML compliance and version control. ‣ Restructured workflows: backlog absorbed, then contractor production capacity increased by up to 50%; designed BI tooling for real-time production tracking.