Principal/Senior Research Scientist, Foundation Models for Credit

ABOUT THE ROLE A new research group is building the next generation of consumer credit scoring from the model architecture up, and this role owns the foundation models at the center of that program. You will design and pre-train encoder-first transformer architectures for heterogeneous tabular and event-sequence credit data, from tokenization to training objective. This is model building from scratch, not adapting off-the-shelf solutions. WHAT YOU'LL DO - Design, implement, and pre-train encoder-first (and encoder-decoder where warranted) transformer architectures for tabular and sequential credit-file data, covering tokenization schemes and training objectives. - Lead the research agenda on representation learning: self-supervised objectives, handling of missingness and censoring, temporal drift, and transfer learning to downstream default, recovery, and account-management tasks. - Own the training infrastructure, including distributed training, experiment tracking, evaluation harnesses, and ablation discipline. - Collaborate with a causal and explainability scientist to make architectures interpretable by construction, using attention structures, bottlenecks, and monotonicity constraints that yield reason codes. - Benchmark rigorously against gradient-boosting incumbents and published tabular foundation models, and write up results for internal, regulatory, and external audiences. WHAT WE'RE LOOKING FOR - 8 to 10 or more years of experience, including 5 or more years building and training transformer encoder architectures from scratch, covering attention mechanisms, positional and temporal encodings, and tokenization of non-text data at scale. - Demonstrated experience pre-training foundation models on millions-plus of examples using distributed training infrastructure (PyTorch Distributed, DeepSpeed, or equivalent). - Prior work in credit, lending, financial services, or another regulated domain where model governance and explainability constrained architecture design. - Strong engineering skills: able to move models from research notebooks to reproducible, deterministic multi-GPU training pipelines independently. - Deep knowledge of tabular and sequence foundation model literature, with a well-reasoned critical perspective on existing approaches. - Experience with self-supervised learning objectives, missing data, censoring, and temporal drift in machine learning systems. - Experience designing interpretable architectures and rigorous evaluation and ablation methodologies. - Experience with event-sequence or time-series transformers (transaction, clinical, or clickstream data) is a strong plus. - Publications at top-tier ML venues (NeurIPS, ICML, ICLR, KDD) or significant open-source contributions to foundation model research are a plus. - Postgraduate research background: a PhD or equivalent research track record preferred. - Must be based in the US; visa sponsorship is not available. LOCATION Hybrid. Open to candidates based in the United States, with presence expected in San Francisco, New York, or Washington, DC.