Deep Learning Researcher

UniversalAGI•San Francisco, CA

2d•Onsite

About The Position

As a Deep Learning Researcher, you'll be architecting and training the foundation models that will transform how industries approach physics simulation and engineering design. This isn't a research role in isolation—you'll be shipping models that customers depend on for critical engineering decisions worth millions of dollars. You'll work directly with the CEO and founding team to push the boundaries of what AI can do with physics data. You'll design novel architectures that can learn from CFD simulations, build training pipelines that scale to petabytes of data, and iterate rapidly based on customer feedback and real-world performance. This is your opportunity to define how foundation models learn physics, from the ground up.

Requirements

3+ years of hands-on experience training deep learning models, with a track record of shipping models to production
Deep expertise in modern deep learning frameworks (PyTorch, JAX) and model architectures (Transformers, Diffusion Models, Graph Neural Networks, GNNs, CNNs, GCNs, PointNet, RegDGCNN, Neural Operators, etc.)
Strong foundation in distributed training: Experience with multi-GPU and multi-node training, gradient accumulation, mixed precision, and optimization techniques
Expert-level Python and proficiency with ML libraries (HuggingFace, PyTorch Lightning, etc.)
Solid understanding of ML fundamentals: Optimization, regularization, generalization, evaluation metrics, and the full training lifecycle
Experience with large-scale datasets: Building data pipelines, handling data quality issues, and working with diverse data formats
Strong intuition for debugging models: Can diagnose training instabilities, convergence issues, and performance bottlenecks
Research mindset with execution focus: Can read and implement papers quickly, but prioritizes shipping working solutions over perfect ones
Outstanding problem-solving: Willing to dive deep into unfamiliar domains (physics, CFD, engineering) and learn what's needed
Excellent communicator: Can explain complex model behavior to customers, engineers, and non-technical stakeholders
Thrives in ambiguity: Comfortable defining what success looks like and figuring out how to get there

Nice To Haves

PhD or Masters in ML/AI, Physics, or related field (or equivalent industry experience)
Published research in top-tier ML conferences (NeurIPS, ICML, ICLR) or physics-ML venues
Experience with physics-informed methods , neural operators, or other physics-ML approaches
Background in physics, computational physics, or engineering (CFD, FEA, multiphysics simulation)
Experience training foundation models or large-scale pretrained models (LLMs, vision models, multimodal models)
Deep knowledge of numerical methods: Quantization, pruning, distillation, efficient architectures
Experience with numerical methods and simulation: Finite element methods, finite difference methods, spectral methods, or other computational approaches to solving PDEs
Experience with geometric deep learning, graph neural networks, or models for 3D data
Built custom CUDA kernels or optimized ML operations for specific domains
Experience at leading AI labs (OpenAI, DeepMind, Anthropic, Meta AI) or high-growth AI startups
Open-source contributions to ML frameworks or well-known model implementations
Forward-deployed experience working directly with customers on model adaptation and deployment

Responsibilities

Design and train foundation models for physics simulation, working with GNNs, CNNs, GCNs, PointNet, RegDGCNN, Neural Operators, transformer architectures, diffusion models, and other cutting-edge approaches adapted for physical systems
Build training pipelines from scratch: data preprocessing, tokenization strategies for physics data, loss functions that capture physical accuracy, and training loops that scale to massive datasets
Optimize model architectures for physics: Balance model capacity, inference speed, and accuracy for industrial use cases with strict performance requirements
Develop novel approaches to physics-informed learning: Integrate physical constraints, conservation laws, and domain knowledge directly into model architectures and training objectives
Fine-tune and adapt models to customer-specific domains, data, and requirements while maintaining generalization and avoiding catastrophic forgetting
Collaborate with infrastructure team to optimize training efficiency, implement distributed training strategies, and ensure models can be served at scale
Validate model performance against ground truth simulations and real-world engineering data, building robust evaluation frameworks that customers trust
Work directly with customers to understand their physics problems, gather domain expertise, and translate engineering requirements into model capabilities
Drive rapid experimentation: Run dozens of training experiments per week, systematically testing hypotheses and improving model performance
Ship models to production: Take responsibility for model quality from initial training through deployment and ongoing monitoring in customer environments