Projects
Investigates how human medical experts interact with LLM-based AI advice in diagnostic settings. Explores expert-AI collaboration where AI serves as an adjunct to human expertise, combining AI's pattern recognition capabilities with human intuition and experience. Focuses on in-the-loop expert engagement in critical medical decision-making, addressing complex factors including patient data integration, variability, and social/legal considerations in healthcare AI adoption.
Implemented a Generative adversarial RL algorithm which is a model free algorithm on the HalfCheetah-v2 gym environment. The core idea consists of extracting expert policy data to yield expected benchmark rewards. Used Trust Region Policy Optimization to overcome the non-monotonicity problem during training.
PyTorch implementation of CycleGAN for unpaired image-to-image translation on the Edges2Shoes dataset. Enables learning mappings between edge sketches and shoe images without paired data. Includes training scripts, evaluation metrics (FID and IS), and preprocessed dataset utilities for experimentation.
Developed a synthetic dataset of 2D images generated from UAV interaction within a Unity3D simulation environment. Implemented domain randomization for augmenting the dataset and improving model robustness. Methods consisted of transfer learning and benchmarking across different image neural nets.
Implementation of YOLOv1 for real-time object detection, trained on a dataset of 10K street-scene images with labeled objects. Includes grid-based predictions, bounding-box decoding, Non-Maximum Suppression (NMS), and a full training/evaluation pipeline for reproducible results.
End-to-end implementations of Region Proposal Network (RPN), Faster R-CNN, and Mask R-CNN for object detection and instance segmentation on a multi-class, multi-object image dataset containing Vehicles, People, and Animals. Includes simplified RPN implementation without FPN, pretrained FPN-backed RPN, and full detector heads for training, evaluation, and visualization on 3,265 images.
End-to-end implementation of SOLO (Segmenting Objects by Locations), a real-time and single-shot instance segmentation model. Detects and segments three types of objects: Vehicles, People, and Animals. Includes data preprocessing, Feature Pyramid Network (FPN) visualization, model training on 3,265 images, and evaluation pipelines with predicted segmentation masks and bounding boxes.