Projects
Things I’ve built and studied
From a patented generative layout system to GPU cluster analysis: mostly AI, with a systems streak.
Content and layout control for generative systems
A granted US patent on controlling both content and layout when generating marketing visuals with generative models. The system I built during my MScAC internship at Ecomtent later became part of it.
Generative AI for product listings and infographics
Designed and built the ML system behind Ecomtent’s AI-generated product images and infographics: diffusion models for imagery, layout-generation networks with constraints on position and size, and stateless API endpoints on AWS.
MLaaS 6.5k-GPU trace analysis
Analyzed scheduling fairness in a 6.5k-GPU production cluster trace. Finding: wait times on low-end, shared GPUs are distributed more fairly than on high-end ones.
Music-score object detection
Applied a Detection Transformer to dense music-notation detection and surpassed the Faster R-CNN baseline.
CT perfusion synthesis for stroke diagnosis
Functional-Consistent CycleGAN for synthesizing CT perfusion maps, plus a UNet Transformer pipeline for stroke lesion segmentation on sparse data, at Sunnybrook’s AICONS lab.
Long-tail video classification
Fine-tuned video classification models on terabytes of imbalanced multimodal data for Huawei Cloud VCT, with full-stack prototypes to visualize real-time inference.