Publications, experience & projects

Center on AI Research for Health (AI4Health), USC Information Sciences Institute · Los Angeles, CA

• Benchmarked multiple state-of-the-art TCR–pMHC binding affinity prediction methods on an in-house dataset to establish reproducible baselines.
• Trained and tuned deep learning models on lab-curated datasets, iterating on data preprocessing, splitting strategy, and evaluation metrics for robust generalization.
• Conducted a literature review of current TCR–pMHC modeling approaches and proposed targeted architectural and objective-function modifications to better fit lab data distributions.
• Designed and prototyped a novel binding affinity prediction approach to advance cancer vaccine research, targeting a NeurIPS submission.

University of California, Riverside · Riverside, CA

• Investigated advanced LLM unlearning techniques to mitigate hallucinations and defend against adversarial attacks.
• Developed evaluation programs to measure LLM performance on diversity, fluency, and utility preservation post-unlearning.
• Used the BLEURT model to assess semantic preservation and identify meaning degradation in unlearned LLM outputs.
• Analyzed academic literature to identify state-of-the-art methods in LLM safety.

Modeling uncertainty in immunology data

Lead project · USC Information Sciences Institute

• Conducting a feasibility study on uncertainty-aware TCR–pMHC prediction under noisy or untested negative labels.
• Investigating cross-reactivity limitations in binary binding models such as TEPCAM, STAG, and TRAP.
• Developing methods to distinguish true non-binding pairs from untested or uncertain TCR–pMHC interactions.
• Evaluating positive-unlabeled learning as a framework for handling uncertain negative labels.

Structure-aware protein interaction modeling

USC Information Sciences Institute · Submitted to NeurIPS 2026

• Co-developed a structure-aware protein interaction model combining ESM-2 embeddings, Cα geometry, persistent homology, and cross-attention.
• Built preprocessing and evaluation pipelines for lab-curated interaction data.

Ensemble learning · 7th place · published in IEEE

Data-science competition (FedCSIS 2025)

• Built a hybrid ensemble for chess-puzzle difficulty prediction using banded base models, stacked meta-learners, and structure-aware features from success curves.
• Designed post-hoc residual calibration to improve high-Elo accuracy and overall error.
• Automated a reproducible pipeline for data cleaning, cross-validated training, ensembling, and submission generation.
• Placed 7th overall; results published in IEEE.

Google Scholar →

Fairness in NLP

Course research project · CSCI 567 (Machine Learning)

• Implemented adversarial debiasing for toxicity detection using BERT representations.
• Designed group-specific threshold tuning and calibration to reduce AAE/SAE false-positive-rate disparities.
• Analyzed fairness–performance tradeoffs across F1, FPR, and FNR under in-domain and transfer settings.

Repository-level code LLMs

Course research project · CSCI 544 (Natural Language Processing)

• Processed RepoBench data and supported the contrastive training setup for repository-level code generation.
• Implemented retrieval baselines and evaluation using Recall@k, Exact Match, Edit Similarity, and CodeBLEU.
• Investigated hallucination-aware context selection beyond similarity-based retrieval for multi-file code completion.