Research on multimodal LLM-based value judgment for audio-visual content, targeting the growing challenge of content governance on short-video platforms. Builds next-gen intelligent content governance algorithms. Key work: (1) daily short-video data monitoring & collection; (2) human-AI collaborative dataset construction for value-oriented fine-grained annotation using Gemini and manual review; (3) compliance analysis of implicit value tendencies in audio-visual content via multimodal LLMs.

Proposed TCRS-QA, a training-free chain-of-thought reasoning framework for shot-aware storyline QA on long videos. Built on Qwen2.5-VL-7B and Qwen2.5-7B, with novel Perception Encoder (PE), Adaptive Keyframe Sampling (AKS), and narrative rhythm detection modules for zero-shot high-quality QA. Achieved 1st place (LLM-Score) in the public >8B track at ICCV 2025 Slamo Workshop SF20K competition.

A multimodal learning-based video description generation system supporting research on multimodal fusion and semantic alignment. Built with Vue3+Flask, supports local upload and URL upload (Bilibili/Tencent Video), handles videos up to 30 minutes, generates real-time captions, and exports WebVTT subtitle files and captioned videos. Collaboration established with Southern New Media.

Supporting system for the national key R&D sub-project 'Camera Position Computation Based on Shot Language Analysis' (No. 2021YFF0900701-3). As a core member, led shot motion recognition algorithm research including classification model design and training; advanced 3D reconstruction algorithms and completed standardized API encapsulation. Built on Vue3+Flask with three-end collaboration (business, algorithm, Unity) for real-time high-precision virtual camera position computation and visualization.

Addressed the high equipment requirements and scalability limitations of traditional detection methods for three DNA methylation types. Based on Nanopore sequencing current perturbation, used GAF to convert hundreds of GB of 1D current signals into 2D images; built EfficientNet, SE-ResNet, and ViT models achieving 88% accuracy (AUC=0.96). Results: National 1st Prize in Life Science Competition (top 0.19%), Zhejiang Provincial Xinmiao Talent Program, ZSTU Key Undergraduate Innovation Program, 1 invention patent pending.

Addressed inefficiency of traditional military knowledge retrieval. Designed three modules: data collection (PDFPlumber+OpenCV+Gaussian filter), extraction & storage (BERT-BiLSTM-CRF/Attention for NER at 93.13% and relation extraction at 93.11%, full pipeline under 5 min), and reasoning (MySQL-Neo4j linkage via Apoc, Cypher queries, D3.js visualization). Won National 3rd Prize at the 14th Service Outsourcing Competition (top 9.091%).

Developed a therapeutic app for high-functioning autism patients using AR robot 'Apiao' as the carrier. Implemented three core modules in Unity/C#: emotion perception (joy/sadness/love), interest cultivation (puzzle/drawing/music/creation), and cognitive imitation. Used ARFoundation SDK for AR scenes, ManoMotion for bare-hand gesture interaction, and facial/emotion recognition to build AR companionship and game incentive mechanisms.

Sole developer. Solved the problem of disorganized literature downloads and vertical-domain knowledge retrieval. Used xlml/requests/selenium to crawl CNKI data (handling iframe nesting), with PyMySQL for data updates. Built full-stack with Vite+Vue3, Element-UI, Echarts, WangEditor, and SpringBoot, supporting conditional search, one-click literature info retrieval, annotation/classification/favorites, auto-categorized downloads, and progress tracking. Awarded Outstanding Result at ZSTU Advisor Project.