AI safety and alignment researcher and Medical Doctor. My work spans alignment, adversarial evaluation and red teaming, uncertainty/calibration, post-training, and the safe deployment of LLM systems in high-stakes settings. I have led end-to-end development of medical LLMs, safety evaluation pipelines, and an Epic-integrated clinical agent used by 3,900+ clinicians across 250,000+ clinical interactions. I bring a rare combination of clinical expertise, research leadership, and deep systems engineering experience across training, inference, and large-scale software infrastructure. I am especially interested in trustworthy reasoning, scalable oversight, and robust deployment of frontier models in real-world environments.

Education

• Ph.D. in Computer Science, Université Catholique de Louvain, Brussels, 2026
  • Defended March 2026
• Doctor of Medicine (M.D.), Université Catholique de Louvain, Brussels, 2023
• M.S. in Computer Science, Université d’Avignon, France, 2014
• B.S. in Computer Science, Université d’Avignon, France, 2012

Experience

AI Safety Research Fellow — Anthropic  ·  Starting November 2026

Independent Researcher — ARISE, Jonathan Chen’s lab, Stanford University (Remote)  ·  August 2026 – present

• Anonymization of clinical data
• Synthetic data generation
• Evaluations of LLMs for medical tasks

Researcher — Sophont  ·  March – May 2026

• Knowledge distillation of reasoning models: compressing chain-of-thought traces before distillation (Compress-Distill)

Research Scientist — Université Catholique de Louvain (Brussels)  ·  October 2023 – March 2026

• Trained Internist-7B, a 7B-parameter medical LLM based on Mistral; first model of its size to surpass 60% on MedQA. Public release: internistai/base-7b-v0.2
• Designed and implemented alignment and safety evaluation pipelines including automated adversarial red teaming, hallucination suppression, self-critique, and metacognitive reasoning assessment for clinical reliability
• Led end-to-end deployment of a clinical agentic system integrated with Epic EHR — model serving, inference orchestration, clinician-facing UI, safety guardrails, auditability, and monitoring; in production with 3,900+ clinical users across 250,000+ clinical interactions. Press: RTL Info
• Visiting Researcher, Harvard University (LiGHT) — evaluated alignment techniques for clinical LLMs with practicing clinicians as part of the MOOVE initiative
• Visiting Researcher, Cleveland Clinic — trained Qwen-based models with GRPO for safe de-identification of clinical documents while preserving semantic fidelity

Machine Learning Scientist (Consultant) — DeepSky (Remote / San Francisco, CA)  ·  August – September 2023

• Designed and curated a dataset for a foundational medical generative text model
• Performed training with AWS SageMaker and PyTorch

Medical Doctor Clerkships — Cliniques Universitaires Saint-Luc (Brussels)  ·  2021 – 2023

• Core rotations: Emergency Medicine, Nephrology, Geriatrics, Obstetrics, Pediatrics, General Surgery, Family Medicine, Pulmonology, Anesthesiology, Radiology
• Medical thesis on computer-assisted diagnosis of rare kidney diseases in emergency departments

Lead Software Engineer — Tilted Phoques (Brussels)  ·  2017 – 2023

• Built large-scale systems in C++, Python, C#, AWS, and Kubernetes, including networking infrastructure handling tens of thousands of real-time concurrent users
• Low-level optimization in Assembly and security analysis using IDA Pro and WinDbg
• Authored open-source Cyber Engine Tweaks — 4,600+ GitHub stars, 10M+ downloads

Software Engineer — Bethesda Softworks (Remote / Austin, TX)  ·  2016 – 2017

• Designed an anti-cheat system from scratch in C++, Assembly, and Python for upcoming titles
• Built backend systems for cheat reports and analytics on AWS using microservices and Lambda; contributed to the open-source AWS C++ SDK
• Researched obfuscation, tamper detection, and code/memory integrity techniques

Software Engineer — ZeniMax Online Studios (Hunt Valley, MD)  ·  2014 – 2016

• Designed load-balancing systems improving response time and capacity at the scale of hundreds of thousands of concurrent connections
• Low-level optimization across memory management, networking, I/O, threading, and lock-free data structures
• Built anti-cheat systems including server-side payload generation, data obfuscation, and debugger traps

Awards & Honors

• 2025 — Senior Area Chair Highlight, ACL 2025
• 2025 — Nature Communications feature
• 2025 — Health Data Agency Grant (€70,000; PI)
• 2025 — FSR Fellowship (Special Research Fund, Belgium) — PhD scholarship
• 2023 – 2027 — FSL Fellowship (Saint-Luc Fund, Belgium) — PhD scholarship
• 2023 — 2nd place, Mistral AI Hackathon (RAISE Summit)

Skills

• Alignment & Safety — RLHF, scalable oversight, preference modeling, red teaming, interpretability workflows, adversarial robustness, uncertainty / calibration
• Model Development — Python, PyTorch, HuggingFace (transformers, TRL), Axolotl, CUDA, C++, Assembly
• Evaluation & Benchmarking — lm-eval-harness, custom clinical reasoning benchmarks, physician-in-the-loop evaluation pipelines
• Infrastructure & Deployment — vLLM, SGLang, Kubernetes, Docker, MLflow, Weights & Biases, Epic EHR integration

Publications

1. Attacking Medical Vision-Language Models with Query-Based Zero-Order Optimization
   M. Griot
   ICML 2026 Workshop Trustworthy AI for Good (AI4GOOD) Workshop @ ICML 2026 2026
   papercite

   Griot, M. (2026). "Attacking Medical Vision-Language Models with Query-Based Zero-Order Optimization." Trustworthy AI for Good (AI4GOOD) Workshop @ ICML 2026.

2. Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation
   M. Griot, P. S. Scotti, and T. M. Abraham
   arXiv arXiv preprint arXiv:2606.05988 2026
   papercite

   Griot, M., Scotti, P. S., & Abraham, T. M. (2026). "Compress-Distill: Reasoning Trace Compression for Efficient Knowledge Distillation." arXiv preprint arXiv:2606.05988.

3. torchtune: PyTorch native post-training library
   M. Obozov, M. Griot, J. Cummings, E. Smothers, F. Mello, R. Ayub, P. J. Bontrager, S. Mohammadi, A. Kwiatkowski, N. Azrak, and M. Mironenco
   arXiv arXiv preprint arXiv:2605.21442 2026
   papercodecite

   Obozov, M., Griot, M., Cummings, J., Smothers, E., Mello, F., Ayub, R., Bontrager, P. J., Mohammadi, S., Kwiatkowski, A., Azrak, N., & Mironenco, M. (2026). "torchtune: PyTorch native post-training library." arXiv preprint arXiv:2605.21442.

4. Medmarks: An Open-Source LLM Benchmark Suite for Medical Tasks
   B. Warner, R. S. Grandhi, M. Kieffer, …, M. Griot, …, P. S. Scotti, and T. M. Abraham (35 authors)
   ICML 2026 Workshop ICML 2026 Workshop on Multi-modal Foundation Models and Large Language Models for Life Sciences 2026
   papercite

   Warner, B., Grandhi, R. S., Kieffer, M., et al., Griot, M., et al., Scotti, P. S., & Abraham, T. M. (2026). "Medmarks: An Open-Source LLM Benchmark Suite for Medical Tasks." ICML 2026 Workshop on Multi-modal Foundation Models and Large Language Models for Life Sciences.

5. Implémentation d'un chatbot dans le dossier patient informatisé
   M. Griot, A. Irrthum, J. Vanderdonckt, and D. Yuksel
   Proceedings Actes de la journée d'étude sur l'utilisation des LLM à l'hôpital 2026
   papercite

   Griot, M., Irrthum, A., Vanderdonckt, J., & Yuksel, D. (2026). "Implémentation d'un chatbot dans le dossier patient informatisé." Actes de la journée d'étude sur l'utilisation des LLM à l'hôpital.

6. MedRiskEval: Medical risk evaluation benchmark of language models, on the importance of user perspectives in healthcare settings
   J.-P. Corbeil, M. Kim, M. Griot, S. Agarwal, A. Sordoni, F. Beaulieu, and P. Vozila
   EACL 2026 Proceedings of the 19th Conference of the European Chapter of the ACL (EACL 2026), Industry Track 2026
   papercite

   Corbeil, J.-P., Kim, M., Griot, M., Agarwal, S., Sordoni, A., Beaulieu, F., & Vozila, P. (2026). "MedRiskEval: Medical risk evaluation benchmark of language models, on the importance of user perspectives in healthcare settings." Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 5: Industry Track), 513–524.

7. A Methodology for Developing and Integrating Large Language Models into Electronic Health Records to Support Clinical Workflows
   M. Griot
   PhD thesis Université Catholique de Louvain — Doctoral Thesis 2026
   papercite

   Griot, M. (2026). "A Methodology for Developing and Integrating Large Language Models into Electronic Health Records to Support Clinical Workflows." Doctoral Thesis, Université Catholique de Louvain.

8. Implementation of large language models in electronic health records
   M. Griot, J. Vanderdonckt, and D. Yuksel
   PLOS Digit Health PLOS Digital Health 2025
   papercite

   Griot, M., Vanderdonckt, J., & Yuksel, D. (2025). "Implementation of large language models in electronic health records." PLOS Digital Health, 4(12), e0001141.

9. A hybrid deployment model for generative artificial intelligence in hospitals
   M. Griot, C. Hemptinne, J. Vanderdonckt, and D. Yuksel
   ML: Health Machine Learning: Health 2025
   papercite

   Griot, M., Hemptinne, C., Vanderdonckt, J., & Yuksel, D. (2025). "A hybrid deployment model for generative artificial intelligence in hospitals." Machine Learning: Health, 1(1), 013001.

10. Pattern recognition or medical knowledge? The problem with multiple-choice questions in medicine
    M. Griot, J. Vanderdonckt, D. Yuksel, and C. Hemptinne
    ACL 2025SAC Highlight Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 2025
    papercite

    Griot, M., Vanderdonckt, J., Yuksel, D., & Hemptinne, C. (2025). "Pattern recognition or medical knowledge? The problem with multiple-choice questions in medicine." Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL).

11. A patient-in-the-loop approach to artificial intelligence in medicine
    M. F. Griot and G. A. Walker
    JAMA Netw Open JAMA Network Open 2025
    papercite

    Griot, M. F., & Walker, G. A. (2025). "A patient-in-the-loop approach to artificial intelligence in medicine." JAMA Network Open, 8(6), e2514460.

12. Physician in the Loop Design of Interactive Agents
    M. Griot, J. Vanderdonckt, D. Yuksel, and C. Hemptinne
    EICS Workshop Engineering Interactive Computer Systems (EICS 2024 International Workshops) 2025
    papercite

    Griot, M., Vanderdonckt, J., Yuksel, D., & Hemptinne, C. (2025). "Physician in the Loop Design of Interactive Agents." In Engineering Interactive Computer Systems. EICS 2024 International Workshops (LNCS 15518, pp. 94–109). Springer.

13. Large language models lack essential metacognition for reliable medical reasoning
    M. Griot, C. Hemptinne, J. Vanderdonckt, and D. Yuksel
    Nat Commun Nature Communications 2025
    papercite

    Griot, M., Hemptinne, C., Vanderdonckt, J., & Yuksel, D. (2025). "Large language models lack essential metacognition for reliable medical reasoning." Nature Communications, 16(1), 642.

14. La régulation de l'utilisation de l'intelligence artificielle en milieu hospitalier
    A. Gobert, M. Rappe, and M. Griot
    Book chapter Droit hospitalier: décodage juridique au départ des réalités hospitalières 2025
    papercite

    Gobert, A., Rappe, M., & Griot, M. (2025). "La régulation de l'utilisation de l'intelligence artificielle en milieu hospitalier." In Droit hospitalier: décodage juridique au départ des réalités hospitalières.

15. Impact of high-quality, mixed-domain data on the performance of medical language models
    M. Griot, C. Hemptinne, J. Vanderdonckt, and D. Yuksel
    JAMIA Journal of the American Medical Informatics Association (JAMIA) 2024
    papercite

    Griot, M., Hemptinne, C., Vanderdonckt, J., & Yuksel, D. (2024). "Impact of high-quality, mixed-domain data on the performance of medical language models." Journal of the American Medical Informatics Association, 31(9), 1875–1883.

16. MetaMedQA benchmark
    M. Griot, C. Hemptinne, J. Vanderdonckt, and D. Yuksel
    Dataset Zenodo 2024
    papercite

    Griot, M., Hemptinne, C., Vanderdonckt, J., & Yuksel, D. (2024). "MetaMedQA benchmark." Zenodo.