Benoit Dherin profile photo

Benoit Dherin

AI Research Scientist at Google

Machine Learning ResearchLLM InferenceMechanics InterpretabilityOptimization DynamicsMathematical PhysicsMLOps Infrastructure

About

I am Benoit Dherin, an AI Research Scientist at Google with a background rooted deeply in mathematical physics and quantization theory. My career has been defined by the bridge between abstract theory and production-ready AI, moving from academic roles at ETH Zurich and UC Berkeley to leading machine learning initiatives at IBM, AT&T, and now Google. I am passionate about mechanics interpretability—specifically understanding how Transformers perform implicit gradient descent—and applying foundation models to critical fields like clinical healthcare. I offer deep expertise in LLM internals and MLOps at scale, and I am always looking to connect with researchers and engineers working at the intersection of foundational research and industry-specific applications to find real mutual value.

Networking

What I can offer

  • Expertise in LLM internals and mechanics interpretability
  • Scaling MLOps infrastructure
  • Coaching technical teams on ML reliability and deployment

Looking for

  • expanding my professional network
  • exploring mutual opportunities in healthcare and foundational AI research

Best fit for

High-level enterprise engineersAI researchersClinical AI collaborators

Current Interests

Implicit gradient descent in TransformersRunge-Kutta methods for deep learningFoundation models in medical imagingQuantum Computing

Background

Career

Transitioned from a decade in academic research and mathematical physics into leadership roles in data science and machine learning at IBM, AT&T, and Google.

Education

PhD in Mathematical Physics from ETH Zürich (2004); Master’s Degree in Mathematics from University of Geneva (1999).

Achievements

  • Published 16+ papers at ICML, ICLR, and NeurIPS
  • Secured US patent for variable learning rate adaptation
  • Proved analytical equivalence between prompt context and parameter updates
  • Developed automated 3D/4D volumetric AI for brain hemorrhages with Mayo Clinic

Opinions

  • Focus on transmuting abstract theory into production-ready, practical AI
  • Deep learning should be stable and interpretable rather than a black box
  • Agile and microservice patterns are essential to prevent reimplementation of models