DeerMe.AI — Camera-Based AI Physical Coaching
Co-founded and led the technical side of a mobile coaching app that reads human keypoints to give real-time form feedback. React Native client, Python vision services, 39 automated tests.
Zihan (Nick) Su 苏梓涵
S.M. student in Data Science at Harvard. I work on multimodal foundation models, video world models, and structured generative modeling.
Open to Summer 2027 research / ML internships
News
- Umbilic Multinomial Logistic Regression accepted to NeurIPS 2026. First author.
- Started at Harvard. Joined Prof. Yilun Du's group to work on video world models.
- Proper Velocity Neural Networks accepted to ICLR 2026. Co-first author.
- Began collaborating with the University of Trento and the Max Planck Institute on geometric deep learning.
Research
I study how structured representations can make multimodal and generative models more capable, from non-Euclidean geometry to action-conditioned video generation.
Recent work covers Flow Matching, hierarchical tokenization, pretrained VLM adaptation, and world models for robot planning.
Publications
Google Scholar ↗-
Umbilic Multinomial Logistic Regression
Extends multinomial logistic regression to umbilic Riemannian geometries, giving one classifier for non-Euclidean representation learning.
Paper and code coming soon
-
Proper Velocity Neural Networks
A full neural-network framework for the unconstrained Proper Velocity model. Best MCC on all five TEB benchmarks, with an 8.3-point gain over HCNN-S on SINEs.
-
Kolmogorov–Arnold Networks-based GRU and LSTM for Loan Default Early Prediction
GRU-KAN and LSTM-KAN combine Kolmogorov–Arnold Networks with recurrent models to flag loan defaults months ahead. Over 92% accuracy three months in advance and 88% eight months in advance, including on out-of-time data.
-
Transforming Credit Risk Analysis: A Time-Series-Driven ResE-BiLSTM Framework for Post-Loan Default Detection
A residual-enhanced encoder BiLSTM that models repayment behavior over time to detect post-loan defaults. Evaluated on 44 cohorts of the Freddie Mac loan-level dataset, with average F1 of 0.92 and AUC of 0.97 against five recurrent and convolutional baselines.
Experience
- Harvard UniversityWith Prof. Yilun Du and Dr. Ruojin Cai · video world models and structured 4D representations.
- University of Trento & Max Planck Institute for Intelligent SystemsWith Ziheng Chen, Prof. Nicu Sebe, and Prof. Bernhard Schölkopf · geometric deep learning and generative modeling.
- Yale UniversityWith Prof. Rex Ying · geometric representation learning for pretrained vision and vision-language models.
Project
Education
- Harvard UniversityS.M. in Data Science · Cambridge, MA
- University of NottinghamB.Sc. (Hons) in Statistics · United Kingdom