Jaskirat Singh Sudan

I'm a dedicated AI researcher and Master's student at the University of Michigan Dearborn, driven to bridge cutting-edge research and practical applications. With a 3.7 GPA in my AI program, I specialize in computer vision, representation learning, and energy-based approaches to extract robust features across modalities.

My academic foundation includes a B.Tech in Computer Science from Medi-Caps University (CGPA: 8.12), where I first developed my passion for data-driven solutions. During my undergraduate studies, I led star-based navigation and RF interference mitigation projects at IIT Indore, securing real-world experience in deep learning and signal processing.

Currently, at UMich TAI Lab, I lead a 5-member team to build an innovative authentication and gesture system using polarized tape patterns on illuminated nails, collaborating closely with Prof. Xiao Zhang.

Connect with me on LinkedIn.

Projects

Contrastive Learning with Siamese Network

Contrastive Learning
Siamese Network

Implemented a Siamese CNN to learn a meaningful distance metric on MNIST, achieving high verification accuracy and smooth latent-space embeddings.

Image Segmentation in Low-Light Conditions

Image Segmentation
Low-Light Driving

Fine-tuned U-Net on low-light driving datasets (A2D2, BDD100K), using advanced augmentation to maintain accuracy under challenging conditions.

Self-Driving Mario Kart

Self-Driving
Mario Kart

Combined CNN+LSTM to capture spatio-temporal features for autonomous control in Mario Kart 64, interfacing via Mupen64Plus for real-time testing.

Speech-to-Image Generation

Speech-to-Image
Latent Diffusion

Fine-tuned a latent diffusion model to generate images from speech prompts, integrating audio feature extraction for multimodal synthesis.

Scalable Convolutional Autoencoder

Scalable Conv
Autoencoder

Developed a hierarchical convolutional autoencoder for large-scale image compression, enabling efficient encoding of high-resolution inputs with minimal loss.

Segment Anything GUI

Segment Anything
GUI

Built a Tkinter GUI for Meta’s Segment-Anything model, enabling intuitive region selection and mask refinement on arbitrary images.

Air-Draw Gesture Interface

Air-Draw
Gesture Interface

Implemented real-time hand-tracking with OpenCV to enable 3D air-drawing, capturing gestures and rendering strokes on screen.

Publications & Research

EEG Artifact Removal Methods

EEG Artifact Removal Methods

Systematic review identifying hybrid ICA+wavelet methods reducing noise by ~30% for real-time BCIs.

ISI Journal, Feb 2024

Online Blogs

Radio Astronomy Blog

Radio Astronomy

Nuclear Conflict Algorithm Blog

Most Important Algorithm That Averted a Nuclear Conflict