Conducted research on vision models at Carnegie Mellon’s Auton Lab, working on model development and evaluation for computer vision tasks.
- Benchmarked a vision transformer (DINOv3) against a VLM (Cosmos) on accuracy and latency for live triage, and containerized a REST-based DINOv3 inference server for integration into the triage pipeline
- Diagnosed a casualty-classification model relying on background shortcuts instead of real injury features using attention visualization, revealing a 12-point accuracy drop caused by an improperly grouped train/test split
- Prototyped a probabilistic model (Bayesian network) in Python that fused multiple observations into injury estimates, validated via Monte Carlo simulation to outperform standalone model output by 13%