Notable Projects

Problem and Importance: NLP classifiers can work well on training-like data but become unreliable when deployed on text from a different domain, platform, or writing style. Overconfident wrong predictions can harm moderation, safety, and decision-support systems.

Method: Developed a post-hoc reliability layer using calibrated confidence, entropy-based uncertainty, and abstain-to-human decisioning.

Results and Contribution: Reduced overconfident errors by up to 13.7%, improved selective risk by up to 49.9%, and maintained a stable 65% automatic acceptance rate on shifted and noisy text datasets.

Problem and Importance: Sentiment models often fail on contrastive language, such as “good, but expensive,” where the final opinion depends on how clauses interact.

Method: Designed a discourse-aware transformer architecture with full-sentence, pre-marker, and post-marker views.

Results and Contribution: Achieved 0.8661, 0.9621, and 0.9071 Macro-F1 on TweetEval, SST-2, and MR.

Problem and Importance: Social media users often express opinions through irony or sarcasm, where literal words do not match the intended sentiment.

Method: Built a multi-task learning framework that learns sentiment, irony, and emotion together using a shared Twitter-RoBERTa encoder.

Results and Contribution: Achieved 0.7241 Macro-F1 on TweetEval Sentiment and 0.869 AUROC for irony transfer with only 0.474% added parameters.

Problem and Importance: Aspect-based sentiment analysis requires a model to identify sentiment toward a specific target, not only the overall sentiment of a sentence.

Method: Developed a dual-graph framework that combines syntactic dependency graphs and semantic co-occurrence graphs.

Results and Contribution: Achieved 0.8994, 0.8763, and 0.8403 Macro-F1 on SemEval-2014 Laptop, SemEval-2014 Restaurant, and SemEval-2016 benchmarks.

Problem and Importance: Cybersecurity teams must prioritize vulnerability reports, but short vulnerability descriptions may not provide enough context for accurate severity classification.

Method: Built a hierarchical classification framework that enriches vulnerability descriptions using Common Weakness Enumeration information.

Results and Contribution: Evaluated 158,379 NVD vulnerability records, enriched 84.3% with CWE context, and achieved 0.7658 Macro-F1.

Problem and Importance: Self-supervised speech models perform well on clean audio but can become unreliable when speech is distorted by noise, reverberation, or enhancement artifacts.

Method: Developed a task-relevant layer disruption analysis framework that compares clean and distorted hidden representations across transformer layers using CKA-based similarity and task-specific layer weights.

Results and Contribution: Built an interpretable diagnostic framework that connects acoustic distortion with task-relevant representation disruption for ASR, speaker verification, and speech emotion recognition.

Problem and Importance: Traffic sensor data often contain missing values, and imputation methods are usually evaluated only by reconstruction error. Low reconstruction error does not always mean better downstream forecasting.

Method: Designed a downstream evaluation study that compares imputation quality using both reconstruction accuracy and forecasting utility.

Results and Contribution: Provided a practical evaluation framework for understanding when traffic data imputation helps or hurts intelligent transportation forecasting tasks.

Problem and Importance: Fine-grained looseness detection often requires dedicated sensors, which limits scalability and increases inspection cost.

Method: Used smartphone accelerometers and vibration-based mechanical modeling to estimate component looseness.

Results and Contribution: Showed how everyday mobile devices can support low-cost structural health monitoring.

Problem and Importance: Deepfake speech can sound natural, and standard acoustic features may miss subtle physical differences in how speech is produced.

Method: Modeled speech through vocal-fold mechanical parameters and used physically motivated features for audio authenticity analysis.

Results and Contribution: Connected speech security with interpretable modeling of the speech production process.

Problem and Importance: Frequent blood glucose monitoring often requires invasive blood sampling, which can be uncomfortable for repeated use.

Method: Developed a machine learning-based framework using ultra-wideband sensing data for non-invasive glucose estimation.

Outcome: Produced an applied biomedical AI framework connecting sensing data with glucose-level prediction.

Problem and Importance: Ransomware attacks encrypt files and block access to important data, especially when users and small organizations lack secure backups or recovery procedures.

Method: Analyzed infection paths, file-encryption behavior, user-level security weaknesses, recovery limitations, and prevention strategies.

Outcome: Developed a practical data protection framework organized around prevention before infection, damage reduction during an attack, and recovery after file encryption.