Projects

Weighted Attention Bottleneck Transformer

Cross-Modal reconstruction, building on Google Research's work on Multimodal Bottleneck Transformer.

Published and presented at University of Toronto ConferenceX'25 (Currently under preparation for submission at CVPR).

A measure theoretic perspective on OOD classification and long-tailed learning

Proposed a unified PAC-learnable anomaly detection framework grounded in measure theory (specifically the Radon–Nikodým derivative), outperforming state-of-the-art methods on 96 real-world datasets.

Submitted to ACM/IMS Journal of Data Science, 2025

Anomaly Detection: Deep Learning and Beyond

Conducted a large-scale benchmark of 20+ ML and DL-based anomaly detection methods across 104 datasets, revealing tree-based methods often outperform deep models in low-data, low-anomaly regimes.

Submitted to IEEE TAI, cited by IBM Research.

Driver behaviour modeling using imitation learning and deep RL

ADAPT (Adaptive Driver Anomaly Perception Technology) is an industry-funded research project focused on quantifying driver behavior unruliness by leveraging human-like traffic data generated through advanced simulation techniques.

High Dimensional Quantile Distribution Shift

Developing robust statistical tools/DL algorithms that are resilient to quantile-based distribution changes.

Submitted to Transactions on Machine Learning Research (TMLR)

Implicit dehazing pipeline for precise image classification

Designed a dehazing-free classification system using pre-trained CNN features and Pinball Twin SVM, achieving 98% accuracy under severe atmospheric noise.

IoT sensor fault detection for smart buildings in Sweden

This was a project in collaboration with Ace-cybersafe(Sweden) where I worked with BrickSchema-modeled IoT sensor data to detect and classify faults using ML techniques. Integrated semantic modeling and real-time data streams for fault monitoring.

BITSAuto: An autonomous vehicle

A funded project for developing an autonomous (self-driving) vehicle.

Led a team of 28 Computer Science undergraduate students and made immense progress on the fronts of: Image Segmentation , Point Cloud Mapping , GNSS Localisation , Path Planning and Controls

Confluence of Epidemiological Modeling and Deep Learning

Presented to Dr. Abhishek Pandey, Associate Director (CIDMA), Yale University and selected for a summer course on Epidemiological Modeling.

Altered Epi-DNNs paper on Indian Covid-19 data and a novel SEIRD model; also presented at a course offered under Dr. Anushaya Mohapatra, BITS Goa