I am an AI Research Engineer at Votee AI and Beever AI Research Lab.
I love mathematics, and I am interested in how mathematical structure can help us better understand machine learning and build AI systems that are more explainable, reliable, and safe.
I also believe that better AI systems can help us make progress in mathematics, science, and other difficult areas of research.
If you are exploring similar ideas, working on related problems, or just interested in learning together, feel free to reach out and say hello!
I study mathematical structure for generalizable AI reasoning. My goal is to understand how geometry, symmetry, constraints, and compositional structure can improve the reasoning, generalization, and interpretability of neural networks. I work on geometric learning dynamics and model adaptation, visual and spatial reasoning, and AI for mathematics and science.
Thoughts on AI, math, and research.
| 2026 Jul | A Local Truth, Globally False: Fable 5 and the Jacobian Conjecture |
| 2026 Jul | How the Jacobian Lens Works: Reading What a Model Is Poised to Say |
| 2026 | What Does Model Growth Really Add? Functional Capacity Beyond Parameter Count. arXiv:2607.14571 [cs.LG], 2026. Preprint. |
| 2025 | MODEST: Multi-Optics Depth-of-Field Stereo Dataset. arXiv:2511.20853v1 [cs.CV], 2025. |
Open source and personal projects.
| SE(3)-Equivariant Transformers | From-scratch implementation of SO(3)-equivariant neural-network primitives — spherical harmonics, Wigner-D matrices, and Clebsch–Gordan tensor products — in pure PyTorch. The geometric building blocks toward an SE(3)-equivariant transformer for multi-view 3D reconstruction. Python, PyTorch, NumPy |
| Code-based Synthetic Multimodal Data Generation | Pipeline that generates synthetic chart-QA data for training Visual Language Models: an LLM writes matplotlib/seaborn plotting code to render charts, while question–answer labels are computed directly from the data with pandas, giving ground-truth answers that are correct by construction. Released a dataset of ~12,600 verifiable QA pairs over ~1,800 charts across three domains on Hugging Face. In a zero-shot baseline, a strong VLM reached only ~40% accuracy (22% on value-reading questions), highlighting a concrete gap in current chart understanding. Built with collaborators for Cohere's Aya Expedition 2025. Python, pandas, Matplotlib, Next.js, TypeScript |