Research at the intersection of AI systems, software security, and machine learning — targeting high-impact venues and real-world applicability.
Existing vulnerability triage relies almost entirely on CVSS base scores, which ignore deployment context and real-world exploitability. This paper introduces the Adjusted Triage Score (ATS) — a contextual risk framework that integrates CVSS authentication sub-metrics, CISA KEV exploit activity data, and EPSS probabilistic exploitation scores. Validated against the NVD REST API v2.0 with a time-stratified train/test split at 2023-01-01, using AUC-ROC and Average Precision as primary metrics and class_weight='balanced' + SMOTE for class imbalance handling.
An open-source framework for holistic LLM evaluation spanning reasoning, instruction-following, creativity, and tool-use tasks. VibeBench includes curated multi-domain datasets, a multi-model evaluation pipeline, and a live leaderboard. Released as v1.0.0 under MIT license.
Leveraging behavioral spending data collected from 70+ students via the Hostel Expense Tracker to train regression and time-series models for monthly expense prediction. Goal: deploy as an in-app module to provide personalized budgeting forecasts.
Status: Data collection & feature engineering phase