Simon Grah

Évaluation des agents RAG — Conception et industrialisation de la chaîne d'évaluation des agents RAG en production : datasets d'évaluation, LLM-as-a-judge en pré-filtre, benchmarking multi-modèles (GPT, Gemini, Mistral) par scoring ELO sur les métriques de tool selection. Passage à l'agentique — Migration d'une architecture RAG classique vers une approche agentic (Google ADK, tool calling) : développement de serveurs MCP exposant le moteur de recherche hybride (sémantique + BM25) aux agents. Boucle de validation experte — Animation de la boucle de validation avec les experts juristes (Label Studio, ateliers, restitutions au PO) et automatisation du pré-filtrage par LLM, réduisant significativement le volume de review manuelle.

• GenAI Production: Industrialized and deployed a secure internal chatbot utilizing Guardrails, serving ~1,000 employees. • RAG Architecture: Designed pipelines for Retrieval-Augmented Generation (RAG) integrated with Knowledge Graphs and Vector Databases to maximize retrieval precision. • Model Optimization: Fine-tuned LLMs and developed advanced NLP solutions on Azure AI. • Technical Leadership: Coordinated multidisciplinary teams and led technological watch on the evolving GenAI ecosystem.

• UK Power Networks (GenAI): Built an end-to-end email processing system using Azure OpenAI and Llama2 for classification, sentiment analysis, and automated response drafting. Developed a RAG chatbot connecting technical documentation to database queries. • Northern Gas Networks: Deployed a predictive maintenance recommendation system for gas network syphons using Neo4j graph modeling, validated by field efficiency gains. • Hutchison Ports: Optimized container flows and truck scheduling via predictive modeling on Databricks.

• OneCBM Platform: Managed the end-to-end ML lifecycle (training, deployment, monitoring) for an industrial FPSO monitoring solution. MLOps Implementation: Established CI/CD pipelines and model tracking using MLFlow. • Data Engineering: Implemented daily batch processing pipelines using Spark for massive dataset handling. • R&D (Data Studio): Delivered 9+ PoCs including predictive maintenance, supply chain optimization, and dynamic pricing.

• Cybels Analytics: Led the technical development of a cyberattack detection platform processing 600 logs/sec via Spark Streaming. • Algorithm Design: Developed graph-based anomaly detection algorithms (DDoS, C2) in Scala and Python. • Research & Open Source: Published research on Model Interpretability (CD-MAKE 2020) and maintained the Shapkit open-source library. • Trusted AI: Conducted research on Explainable AI (XAI), Fairness, and robustness against adversarial attacks for European defense projects (Artemis, H2020).