Developed production-ready computer vision models for object detection and segmentation in a Point Ambiguity Resolution system, focusing on spatial grounding, multi-object disambiguation, and accuracy–latency tradeoffs for real-time use. Built and deployed end-to-end ML pipelines for real-time inference and batch processing, including 3D scene reconstruction and Gaussian Splatting–based representations, optimizing GPU inference to achieve 1.8 ms latency using TensorRT. Performed benchmarking and ablation studies across perception backbones and Vision–Language Models (SAM2, DINOv3, YOLOv11n-seg), achieving 38.9 mAP (detection) and 32.0 mAP (segmentation) with tracking in Weights & Biases. Designed and evaluated diffusion and Flow Matching–based generative models for trajectory and endpoint prediction, conditioned on scene context and joint state, using probabilistic modeling and controlled ablations.

Graduate Research Assistant at UMN Computer Networking Lab, University of Minnesota

August 1, 2024 - August 1, 2025

Engineered scalable data processing and feature engineering pipelines for large-scale time-series datasets (200+ hours), supporting regression, classification, and forecasting with automated data validation and quality checks. Trained and evaluated supervised ML models including linear and tree-based regression, Temporal Convolutional Networks (TCN), XGBoost, Random Forest, Deep Neural Networks, and LLM-based predictors, achieving 85%+ prediction accuracy. Implemented temporal feature extraction and systematic model evaluation, improving predictive performance by 23% over baseline using cross-validation, hyperparameter optimization, and standard metrics (MAE, RMSE, R^2).