CĂ©dric Caruzzo
AI Research Scientist at Lunit Inc.
Vision & Language Foundation Models · Post-Training · Safety & Alignment
Vision & Language Foundation Models · Post-Training · Safety & Alignment
I am an AI Research Scientist at Lunit Inc., where I work across the full AI stack for precision oncology — from vision and language foundation models through post-training, safety, and alignment. My role spans both research and engineering: designing the systems and training pipelines that make large-scale biomedical AI reliable, interpretable, and deployable in clinical settings.
On the vision side, I train self-supervised ViT-based pathology foundation models (DINOv2/v3-style) and build the distributed inference infrastructure to run them over hundreds of terabytes of whole-slide images, reducing compute overhead by 60–80%. In parallel, as part of a government-led national consortium (Chain of Evidence), I contribute to building medical LLMs from scratch — curating 292B+ token corpora, designing RAG pipelines and knowledge graphs, post-training and aligning 21B dense and 16B MoE models, and developing rigorous evaluation frameworks for medical reasoning.
A significant focus of my work is AI safety and alignment in deployed medical models — studying failure modes, developing safety guardrails, and applying mechanistic interpretability to identify and correct problematic circuits and attention behaviors. I also build multi-agent agentic frameworks for clinical reasoning over structured and unstructured biomedical data.
I hold a Master's degree from the KAIST Graduate School of AI (KAIST Global Presidential Scholar), where I was advised by Prof. Ye Jong Chul in the BISPL lab, focusing on foundation models for Cell Painting assays. Prior to KAIST, I was a Data Scientist in the Pasteur International Unit AI3D (joint initiative between Institut Pasteur Paris and Institut Pasteur Korea), where I built the core computer vision pipelines for high-content screening and drug discovery.