I'm Khushal, a first-year CS Master's student at Northeastern's Khoury College of Computer Sciences, focused on computer vision and applied ML. Before this I did my undergrad at Ahmedabad University in India, where I co-built the Padik dataset on pedestrian-driver interactions for ADAS, published at IEEE ICVES 2024. After that I spent few months at ISRO's Space Applications Centre fine-tuning YOLOv8 on Mars satellite imagery for sand dune detection, accepted at the 8th International Planetary Dunes Workshop. These days I'm working on attention-based vision models, parameter-efficient fine-tuning, and RAG systems.

Curious about video understanding at scale, parameter-efficient adaptation of large vision models, multimodal systems combining vision and language, and the messy reality of deploying ML on edge devices. Looking to connect with people working on applied research at the perception layer of physical AI, autonomous systems, or smart devices, especially folks who care about the gap between benchmark numbers and what actually works in production.

Right now I'm tinkering on an attention-based gaze estimation system using DeiT-Small with rank-4 LoRA adapters. Personalizes to a new user from 9 calibration samples by training only 0.56% of the model parameters, cuts mean angular error from 7.96° to 5.56° on MPIIFaceGaze, and runs live in a webcam demo with MediaPipe face normalization. Also poking at video data pipelines and parameter-efficient fine-tuning for vision transformers more broadly.