At the University of Toronto’s BMO Lab, I built a real-time Python and PyTorch pipeline that streamed captured human movement into humanoid control inputs for physics simulations. I integrated timestep normalization, noise filtering, and video-based body calibration, reducing motion-tracking error by 24% and fall-down failures by 26%.
I also extended a human-pose model with preprocessing, prediction heads, and loss functions, then trained four torque-prediction variants while keeping pose-reconstruction error within 2% of baseline.
At CIBC, I designed IBM MQ integration for EquiLend that replaced 15-minute NAS transfers with near-real-time messaging. I also improved ELT reporting availability by 30% through incremental loads and parallel execution, and supported production applications processing millions of records daily.
In my RFP Qualification Agent project, I developed an LLM-powered backend that produced traceable bid/no-bid recommendations from large PDF documents, using PostgreSQL, pgvector, and FastAPI. For RecipePrep, I led technical development of a generative AI system that used specialized LLM agents, RAG, and recipe and nutrition data to generate recipes around user constraints.
