Projects

Optimized Qwen3-VL-2B inference on Snapdragon 8 Elite, achieving a 2.05× speedup over the project baseline. Configured llama.cpp with importance-matrix-based Q4_0 quantization, memory locking (mlock), and CPU affinity for on-device inference. Fine-tuned with LoRA for visual question answering, evaluating both task quality and inference performance.

Explored a variety of model architectures for EEG signal analysis, including CNN, RNN attention-based models, Transformers, and hybrid models.

Embarked on evaluating and enhancing the performance of Large Language Models (LLMs) in detecting fairness and factuality in textual claims.

Leveraged DiffEdit, an innovative text-conditioned diffusion model for semantic image editing, integrating it with BLIP and other models into an interactive framework for a seamless end-to-end image generation and editing pipeline. Additionally, pioneered a novel technique for text-guided mask generation in DiffEdit, enabling precise object segmentation through textual queries.