Document Preprocessing and Structuring Pipeline (Undergraduate Research at FIB)
Built an industrial-grade PDF-to-JSON pipeline using modular layout analysis and LLMs to structure complex Chinese technical documents with SOTA accuracy.
Built an industrial-grade PDF-to-JSON pipeline using modular layout analysis and LLMs to structure complex Chinese technical documents with SOTA accuracy.
Engineered a 5-stage pipelined MIPS CPU on FPGA to accelerate sparse matrix multiplication, validating hardware correctness through real-time visualization.
Authored a comprehensive technical manuscript for Probability and Stochastic Process (I), structuring concepts from measure theory to entropy with rigorous proofs and custom visualizations.
Proposed a supervision-free alignment framework for LLaVA 1.5 that utilizes a Reward Union strategy to reduce object hallucination by 81.2%, outperforming GPT-4V.
Developed a full-link communication simulator integrating Viterbi decoding and custom encryption to optimize the trade-off between transmission security and image recovery quality.
Engineered a C++/Qt6 client-server vehicle management system featuring O(log n) data retrieval, SHA-256 security, and cross-platform compatibility verified by GoogleTest.