What you will do

• Develop methods to map latent embeddings to human concepts, with the explicit goal of labeling embedding dimensions/features as interpretable text descriptions relevant to credit risk modeling.
• Design and execute mechanistic interpretability research on Transformer architectures (e.g., probing representations, causal interventions, and internal component analysis) to understand how credit risk signals are encoded.
• Build an end-to-end interpretability workflow: experimental design, implementation, evaluation, and clear documentation of findings and limitations.
• Engineer research code into durable tooling: refactor experimental notebooks/prototypes into clean, modular, testable Python code that supports iteration and reuse.
• Collaborate with credit risk domain experts to ensure interpretability outputs are meaningful, actionable, and grounded in domain reality.
• Partner with internal MLOps / ML engineers to run large-scale training and integrate research tooling with GPU/cloud execution environments—while keeping momentum when support is intermittent.
• Contribute to technical strategy: propose experiments, define success criteria, and help de-risk the approach through principled iteration and evidence.

What experience you need

• Education & Experience: A PhD in a quantitative discipline is highly preferred (with a minimum of 5+ years of relevant professional/research experience) OR an MS with 7+ years of exceptional, specialized eXplainable AI (XAI) experience is acceptable.
• A strong, demonstrable background in interpretability, representation analysis, or research/tooling focused on Transformer models.
• Strong mathematical and statistical foundations (linear algebra, statistics) sufficient to implement/interpret representation analysis methods (the work references advanced techniques such as Kernel CCA).
• Experience with LLMs and Transformer development workflows (e.g., common training/evaluation patterns).
• Communication & documentation strength: ability to clearly explain complex findings to technical stakeholders and produce crisp experimental notes and interpretability artifacts.

What could set you apart

• Strong research engineering skills in Python with demonstrated ability to implement and iterate quickly while maintaining code quality and reproducibility.
• Expertise with modern deep learning frameworks (TensorFlow strongly preferred / PyTorch acceptable).
• Hands-on experience in Mechanistic Interpretability / XAI for Transformers, including the ability to run controlled experiments that isolate and explain internal model behavior (e.g., activation-based interventions and interpretability analysis patterns).
• ML engineering experience running Transformers on cloud GPU infrastructure, including practical understanding of training large models with NVIDIA GPUs and troubleshooting training/runtime issues.
• Familiarity with Hugging Face transformers/datasets ecosystems and applied workflows for training/fine-tuning and data iteration.
• Exposure to high-volume data handling for time-series or sequential modeling (e.g., efficient data loading strategies; pipeline integration patterns), especially when datasets are large and irregular.
• Experience applying interpretability techniques to regulated or high-stakes domains (finance, healthcare, compliance-heavy environments), where explanation quality and defensibility matter.
• Evidence of impact via publications, open-source contributions, or internal tooling related to interpretability, representation learning, or Transformer analysis.

Why this role is different

• You’ll tackle a high-novelty interpretability problem: translating Transformer embedding spaces into human concepts for real-world credit risk time-series data.
• You’ll have access to domain experts for credit risk meaning and labeling support—so your focus stays on the hardest technical bottleneck: mechanistic interpretability.

LI-AM2

LI-Hybrid