Cyril G.

AI-powered railway data platform

• LLM assistant built on a self-hosted open-source model, combining a custom tool-orchestration layer with RAG over internal documentation (pgvector), letting users query logs, trigger analyses, and generate dashboards conversationally.
• Distributed processing pipeline with containerized workers orchestrating data ingestion, parsing, and analysis.
• Interactive investigation tool correlating multiple data sources through time-based replay and topology visualization, surfacing insights not visible when analyzing each source in isolation.

• Model development: Designed, trained, and optimized deep learning models end-to-end, from architecture and training to deployment, including human pose estimation and action recognition running in real time on embedded hardware (NVIDIA Jetson).
• Real-time inference pipelines: Built multi-stream inference pipelines for embedded systems, optimizing models (ONNX, TensorRT) for efficient on-device deployment under tight resource constraints.
• Training infrastructure: Developed a custom orchestrator to schedule and manage distributed model training jobs across a fleet of on-premise NVIDIA DGX Stations, with experiment tracking via MLflow.
• Data & annotation tooling: Defined data pipeline requirements and built internal tooling for dataset management and annotation workflows; established annotation guidelines and quality standards, coordinating with annotation teams to ensure consistent, high-quality labeled datasets across projects.