• Built an automated evaluation framework for multi-turn LLM conversations, standardizing quality and safety metrics across 5M+ customer–LLM dialogues/day; defined a failure-mode taxonomy and continuous regression test suite to score question quality, logical flow, hallucination, and policy compliance, and detect regressions after model/prompt updates.

• Built an end-to-end evaluation pipeline for customer comments and LLM-generated follow-up questions, eliminating manual review and saving $1M+ and 25,000+ labeling hours/year; used RAG-based retrieval and prompt optimization to quantify information gain, relevance, and clarity, with synthetic edge-case generation to expand coverage on hard scenarios.

• Designed and deployed production Responsible AI guardrails for return-related chat, sustaining 50+ TPS real-time inference; implemented prompt-injection defenses, misuse/content-violation detection, and risk-based routing, optimized for low-latency and cost-efficient serving under adversarial traffic patterns.

• Extended diffusion-model safety into an image-safeguarding framework that protects customer images from generative misuse within Amazon’s Responsible AI ecosystem; published at WACV 2025, with a pending patent and the Amazon Inventor Award for innovation and impact.

• Developed a novel and practical defense method to protect user images in real-time from identity theft and copyright misuse in generative diffusion models, demonstrating real-time protection applicability, by introducing latent-space adversarial perturbations.

• Engineered an optimized perturbation generation and inference framework, accelerating processing speed by 150× (from ~30s to 0.2s per image) and reducing memory usage by 2.6×, by optimizing the perturbation generation path while maintaining protection performance.

• Evaluated the system on three public diffusion benchmarks and 10,000+ real-world samples, showing up to 20% improvement over existing diffusion-model protection methods across 10+ evaluation metrics, by measuring protection strength, visual fidelity, and generalization end-to-end.

• Developed an end-to-end machine learning pipeline for data cleaning, labeling, and DNN training, powering an in-house ID verification system used by 10M+ users, by enabling accurate ID card information extraction.

• Designed and deployed a secure internal data-labeling platform integrated into the pipeline, processing 20,000+ high-quality annotations and reducing external labeling costs by more than $100K, by moving labeling workflows in-house with secure, integrated tooling.

• Enhanced OCR model performance, improving accuracy by 10% in large-scale deployment, by targeted data curation and model fine-tuning to deliver production-level reliability.

• Conducted leading-edge research in Deep Learning, expanding research direction and idea exchange, by identifying emerging research challenges/opportunities and hosting a weekly paper reading group.

• Designed and executed extensive experiments to validate hypotheses, resulting in 20 published papers at top AI conferences (NeurIPS, ECCV, AAAI, ICASSP, MobiCom), by effectively communicating research findings through concise academic papers and conference presentations.

• Supported course instructors, mentoring and assisting over 200 students in foundational and advanced Computer Engineering courses, by organizing weekly recitations, grading assignments, and holding office hours.

• Installed, upgraded, maintained, and ensured smooth operation of over 100 PCs, Macs, printers, projectors, and mobile devices.

• Provided top-notch customer support and resolved client issues in person, via phone, and remote desktop connections for more than 100 staff and faculties with a wide range of technical backgrounds.

• Performed diagnostics and troubleshooting system issues, documented help desk tickets/resolutions, and maintained equipment inventory lists.

• Planned a 7-day-long orientation program for more than 2000 international students.

• Helped more than 1000 new students integrate into American academic culture via social events.

• [CVPR-23, 24] IEEE / CVF Computer Vision and Pattern Recognition Conference

• [ECCV-22, 24] European Conference on Computer Vision

• [NeurIPS 23, 24] Conference on Neural Information Processing Systems.

• [ISCAS-23] IEEE International Symposium on Circuits and Systems

• [AICAS-22] IEEE International Conference on Artificial Intelligence Circuits and Systems

Rutgers University, New Brunswick, NJ, USA

• [Fall 2022] Digital Logic Design

• [Spring 2022] Probability and Random Processes

• [Spring 2024] Introduction to Deep Learning