I work on understanding and steering the internal representations of neural models.
My main interests are:
- mechanistic interpretability and representation engineering in LLMs
- probing, directions, and geometry of representation spaces
- control-oriented and representation-driven modeling in computational genomics (e.g., breast cancer networks from cBioPortal)
Recent projects include white-box sycophancy detection in LLMs, multi-turn retrieval-augmented generation systems, and control-ready gene regulatory networks for breast cancer.
Technical stack
- Languages: Python, C/C++, Bash
- ML/Systems: PyTorch, HuggingFace, FAISS, scikit-learn, NumPy
- Infra: Linux, SLURM/HPC, Git, Docker
Some personal interests
- Philosophy of explanation and what it means to understand optimized systems
- Geometry of representation spaces in high dimensions
- Learning new sports and small adventures
- Long walks, bad coffee, and occasionally recompiling kernels for no good reason