I am a Senior AI Research Scientist specializing in Physical AI, Self-Supervised Learning, and robust vision encoders. Formerly pioneering research at UC Berkeley (BAIR) and deploying models to millions at Microsoft Research.
Architecting next-generation vision models at Nota AI. Designing video-distilled single-image encoders using Diffusion Transformers (DiT) to understand real-world physics without brittle external trackers.
Foundation models deployed in dynamic domains suffer from critical reliability failures. We propose a lightweight paradigm that distills temporal knowledge from video into a standard single-image encoder, setting a new SOTA for DINO-style video distillation.
Read Abstract / Paper
Most SSL methods train on static images, missing temporal cues. We introduce a video-distilled single-image encoder trained to predict the next-frame representation, injecting 3D spatial and temporal priors without optical flow.
Read Abstract / Paper
Consolidated PhD findings. Demonstrated that "implicit" pooling methods developed at Berkeley outperform traditional "explicit" part models. Introduced 'Activation Flow' visualization.
Invented "Generalized Orderless Pooling" (Alpha-Pooling) during time as Visiting Research Scholar at UC Berkeley (ICSI / BAIR) with Prof. Trevor Darrell.
Pioneered learning part models in a completely unsupervised manner, without part annotations or bounding boxes, achieving SOTA on fine-grained recognition datasets.
Translating academic background into production-ready AI models. Writing clean, modular code used globally.
Trained and integrated an image classification model for animal species recognition directly into the Bing search engine. Powered the "Species Recognition" feature, successfully handling millions of daily inferences in production.
Core contributor to Microsoft AI for Earth's `CameraTraps` and `SpeciesClassification` platforms. Developed tools allowing conservationists globally to run MegaDetector and DeepFaune models.