Foundation models for scientific and industrial data.
Expert AI: foundation models for the data that runs science and industry
We define Expert AI as foundation models built for scientific and industrial data: continuous and irregular time signals, spatiotemporal fields, multimodal scientific recordings, time-stamped event logs from real organizations. The kind of data that doesn’t look like text, doesn’t fit transformer assumptions, and rarely leaves the institutions that hold it. This isn’t a fine-tuning problem. It requires architectures designed for the data, infrastructure designed for how the data lives, and a tech stack of AI components purpose-built to assemble foundation models for new domains.
General-purpose AI was never built for science and industry. The data is sparse, noisy, distributed, and governed by strict rules. It demands expert intelligence: models that understand the structure of a domain, not just the surface of its data.
Our approach is co-foundation with domain experts: AI and domain science evolving together. We build domain-specialised foundation models trained to master a field’s data, constraints, and logic from the ground up. We work hand in hand with the scientists, clinicians, and engineers who know the domain best. Our models learn from distributed sources through decentralised architectures, keeping sensitive data where it belongs.
We anchor in brain and neuroimaging, the most complex signal of all and the hardest class of expert data we know how to characterize. In parallel, our cross-domain transfer programme works on extending the same components and methods to open new application domains.
We invest in demonstrators that turn research into tangible proof, working with partners on concrete use cases that test our models in real conditions and inform the next iteration of the science.
Our team finished 2nd out of 1,183 in the NeurIPS 2025 EEG Challenge, and we publish at Tier-1 AI conferencesincluding NeurIPS, ICML, and CVPR. This dual focus, rigorous science and applied impact, defines our approach.
Sigma Nova is looking for a research intern (starting in Q1 / early Q2 2027) to investigate fundamental questions in machine learning: how can models adapt to new datasets, express reliable uncertainty, and learn from heterogeneous sources while respecting privacy?
This internship combines mathematical thinking with hands-on experimentation. You will work with researchers on methods that connect modern foundation models with ideas from probability, Bayesian inference, and optimal transport. The exact topic will be shaped around your interests and background, with a focused scope and opportunities to contribute to a publication.
1) Bayesian learning and uncertainty
Explore how models can provide useful uncertainty estimates alongside their predictions. Possible directions include scalable Bayesian approaches, parameter-efficient adaptation, and evaluating reliability when models encounter unfamiliar data.
2) Learning from distributed and heterogeneous data
Investigate how models can learn from multiple data sources without bringing all the data together. Topics may include federated learning, combining information across different populations, and understanding trade-offs between predictive performance, communication costs, and privacy.
3) Optimal transport and adaptation
Study how the geometry of probability distributions can help models compare datasets and transfer knowledge between tasks. Possible directions include learning compact dataset representations, selecting informative examples, and improving adaptation from limited observations.
These are alternative starting points rather than a single combined project. Depending on the topic, the balance may lean toward method development, mathematical analysis, or empirical evaluation.
Master’s-level training in applied mathematics, statistics, machine learning, or a related field
Strong Python skills and experience with PyTorch
Solid foundations in probability, linear algebra, and optimization
Comfortable reading research papers and designing careful, reproducible experiments
Interest in Bayesian inference, optimal transport, or privacy-preserving learning
Prior experience in one of these areas is a plus; expertise in all of them is not expected
A research-driven startup working on foundation models
Close mentorship from researchers working on optimal transport and Bayesian deep learning
High-end GPU resources and existing research code to build on
Room to shape the research direction and contribute to a publication
1) Prescreen with Recruiter
2) Hiring meeting with one NeuroAI team member
3) Onsite interview (coding and research discussion)