AI-Powered Assistive Wearable for the Visually Impaired

A low-cost, Raspberry Pi–based wearable that combines object detection, depth estimation, OCR, and multimodal feedback to support navigation and reading for visually impaired individuals in unstructured environments.

Features: (1) YOLO-based object detection with monocular depth for obstacle identification and distance estimation with audio/vibration alerts, and (2) OCR pipeline (PyTesseract + TTS) for reading text from signage.

• Published in American Journal of Computer Science and Technology
• Patent granted by Intellectual Property India

Low-Cost Wearable EEG System for Real-Time Cognitive State Classification

-> A sub-$35 wearable EEG system capturing raw cognitive signals at 80–100 Hz, cutting hardware costs by over 94% compared to standard commercial-grade devices.

-> Curated a dataset of 400 seconds from 10 participants, utilizing standard serial communication to log raw analog-to-digital signals directly without applying hardware offset corrections
-> Evaluated five machine learning models on raw data to achieve 56% test accuracy, demonstrating a measurable 6-point improvement over random chance with 0.55 ROC-AUC values
-> Demonstrated practical feasibility by integrating the predictive models into an online car simulator, utilizing classified attention and relaxation states to control acceleration and braking dynamically.

• Published in Journal of Engineering Research (Q1, Elsevier)

BIPV & Floating Solar Solution for Scalable Energy Infrastructure

Selected as a Student Researcher for the New York Academy of Sciences — Junior Academy, a highly competitive global program with roughly a 10% acceptance rate.

Collaborated with a team of 6 high schoolers from across the world and developed a BIPV and floating solar based solution for the probelm statement " To design an innovative and scalable solution to improve electrical infrastructure and/or energy storage technology in order to make solar energy use more reliable, efficient, and economical for meeting the energy demands of technology and society."

• Student Researcher, New York Academy of Sciences — Junior Academy (~10% acceptance rate)

Image Segmentation for identifying Brain Tumors in patients' MRI scans

Developed an anatomically guided 2D U-Net with prior-informed attention for brain tumor segmentation (glioma, meningioma, pituitary) using the BRISC MRI dataset in PyTorch, achieving a Dice score of 0.709 and HD95 of 36.5 px. Engineered a complete training and evaluation pipeline incorporating class-imbalance-aware loss functions, 3-fold cross-validation, geometric data augmentation, and multiple segmentation metrics.

Benchmarked the model against baseline U-Net, attention-gate ablation, Swin-UNet, and UNETR architectures, using paired Wilcoxon tests to assess significance. Conducted false-positive analysis to evaluate anatomically implausible detections and compared model performance under identical training conditions.