ls -l
  • GraphHands/ currently working on A graph-based pipeline that learns a shared latent space between human and different robot hands for cross-embodiment manipulation.
  • Vision-Based Drone Flight/ A quadrotor that follows another drone using only its onboard camera — a YOLOv5 detector trained on synthetic renders feeds a PPO tracking policy learned in simulation and deployed on a real drone.
  • Medical RAG Assistant/ A privacy-first clinical decision support system — manage patient records, transcribe consultation audio, and answer free-text medical questions with cited, RAG-grounded answers, all running fully locally so no patient data leaves the machine.
  • Monocular Visual Odometry/ A robust monocular visual odometry pipeline that estimates camera trajectory from a single video stream — combining Essential-Matrix initialization, bidirectional KLT tracking, two-stage PnP-RANSAC pose estimation, and continuous landmark triangulation, tuned to run on KITTI, Malaga, Parking, and a self-recorded handheld sequence.
  • MRI PoPE-ViT/ A Vision Transformer pipeline for brain-tumor MRI classification that pits Polar (PoPE) against Rotary (RoPE) positional embeddings to test whether explicitly decoupling 'what' from 'where' improves diagnosis — benchmarked against pretrained and convolutional baselines.
  • Action Persistence in Deep RL/ A study of what happens when a reinforcement-learning agent holds each action for k steps — fixed and learned persistence for PPO, SAC and DQN, compared on reward, sample efficiency, control smoothness and robustness to action delay.