M.S. in Artificial Intelligence (’26) || University of Michigan–Dearborn
M.S. student in AI at UM-Dearborn. At ISSF Lab, my Master’s thesis studies contrastive representation learning for robust deepfake detection, focusing on how loss design and similarity choices shape embedding geometry and out-of-domain generalization. Earlier in the lab, I implemented a from-scratch replication of a two-stage deepfake detection framework and built speaker-specific diagnostics for analysis. Before this at TAI Lab, I lead a 5-member team developing a visible-light tag authentication system. Previously a Research Assistant at IIT Indore (star-catalog navigation, RF-interference mitigation for GMRT). B.Tech in Computer Science, Medi-Caps University (’24). My interests are computer vision and machine learning, especially representation learning and self-supervised embeddings.
I believe real understanding can’t come from text alone. Like humans, models need perception to build “mental models” of the world. My work uses representation learning and self-supervised learning, pretext tasks like mask-and-reconstruct for images, video, and audio to force models to form rich embeddings (their internal mental images). By shaping these embeddings with tailored losses, I steer what models attend to and how they reason. Goal: systems that truly grasp the world, not just the words about it.
Investigates how supervised contrastive learning regularizes latent representations from self-supervised speech encoders to improve deepfake audio detection robustness and out-of-domain generalization.
Implemented a from-scratch replication of a two-stage deepfake detection framework: Stage-1 learns compressed style/linguistic representations from SSL encoders and Stage-2 trains a classifier with augmentation. Includes Wav2Vec2/WavLM variants, EER evaluation, and speaker-specific embedding/diagnostic visualizations.
Built a Siamese network on MNIST with contrastive loss to learn a 128-D embedding where similar digits are close and dissimilar ones far apart; includes a Tkinter+Plotly GUI for pairwise similarity, few-shot classification, and 3D embedding visualization.
End-to-end agent using screen capture + CNN-LSTM for temporal understanding, controlling the Mupen64Plus emulator via keyboard events. Repo covers data collection, training, and autonomous play pipelines with usage instructions across OSes.
Evaluated MobileNetV2-U-Net and Xception-U-Net on BDD100K with transfer learning and targeted fine-tuning for nighttime scenes; README reports metrics (e.g., Xception-U-Net Dice ≈0.90, strong low-light gains) and outlines preprocessing, losses (Dice/BCE), and results.
Pipeline that transcribes speech to text via Whisper and generates images using Stable Diffusion v2 fine-tuned with DreamBooth; README details motivation, dataset creation, training strategy (low LR, mixed precision), and links to presentation/demo.
Lightweight Tkinter interface for Meta’s SAM: load checkpoint (ViT-H/L/B), add point/box/text prompts, and visualize segmentation with a progress indicator. Usage instructions and model download links are included.
Trainable local autoencoder with a simple GUI to watch reconstructions evolve from the latent space; repo includes runnable scripts, requirements, and quick-start steps.
Estimated latitude and longitude of the object by matching sky/constellation patterns against star catalog references. Explored multiple matching strategies (template matching, feature matching + homography, and a grid-based approach) and combined methods to improve rotation/scale robustness.
Real-time virtual whiteboard for online classes/meetings. Draw, erase, and change color/thickness using simple hand-gesture controls, enabling live annotation with a webcam-based interface.
How radio telescopes see what eyes can’t, revealing pulsars, black holes, and interstellar gas to map the invisible universe.
How an early-warning algorithm and crucial human judgment prevented a Cold War false alarm from escalating into catastrophe.
