Graybox is an early-stage MLOps startup building a model training solution for deep neural networks.
Our goal is to create a platform that simplifies the experimentation process, making it easier, faster, and more cost-effective for any AI developer.
We are an international team split between Paris and Zurich.
We’re looking for a PhD student with a strong background in Machine Learning to explore new training strategies for building leaner, more interpretable AI Models (neural networks). You'll work with us on Graybox, a research toolchain designed for fine-grained intervention in model training — including neuron freezing, pruning, and progressive expansion.
This opportunity is ideal for someone who enjoys blending theory with experimental practice, and wants to help define more principled ways to grow and structure models — not just train them blindly.
Explore progressive model growth strategies: starting from minimal architectures and selectively expanding capacity during training
Experiment with controlled neuron activation, using curated input subsets to guide learning
Design and test pruning or slimming strategies that preserve performance while improving interpretability and efficiency
Analyze redundancy and generalization in learned representations
Deliver short internal write-ups or summaries that capture your findings
Are currently pursuing a PhD in Machine learning, Computer Science, AI, or a related field
Have strong implementation skills with PyTorch or other DL frameworks
Are curious about model structure, training dynamics, or explainability
Are excited to try unconventional approaches like guided neuron emergence, progressive network assembly, or active pruning
Prefer thoughtful, open-ended research over benchmark chasing
Experience with transformers, foundation models, or efficient model architectures
Familiarity with Dash, gRPC, or ML ops tools
Interest in interpretability frameworks, visualizations, or open-source ML infrastructure
A focused, flexible internship working on deep training mechanics and model transparency
Hands-on mentorship and room to explore novel research directions
Potential to publish, contribute to open-source, or shape future research tooling
A chance to influence how foundation models are made more modular, efficient, and explainable
