Built a privacy-preserving on-chain lending prototype for real-estate collateral, validating property verification, LTV compliance, and loan state management.

• Implemented a zero-knowledge compliance flow for LTV≤80%, proving loan eligibility without publishing property valuation or loan principal in plaintext.
• Built core modules for property verification, compliance proof, loan request, approval, cancellation, and repayment, with a property-level active loan lock to prevent duplicate collateralization.

Built skeleton-sequence and Transformer-attention models for classifying and scoring Baduanjin and Wuqinxi movement videos, winning First Prize in the Northeast Regional AI Technology Track.

• Used Transformer Encoder / multi-head self-attention structures to classify 15 martial arts action categories and predict movement-quality scores, outputting action labels, scores, and inference latency.
• Won First Prize in the Northeast Regional AI Technology Track and received an Outstanding Award at the National Finals hosted at Nankai University.

Contributed to internal efficiency tooling in a bank IT environment, developing and maintaining VBA automation programs for report processing, repetitive workflow automation, and reliable data flow.

• Optimized code structure, exception handling, and maintenance logic for VBA automation tools, reducing manual error risks and improving maintainability.
• Participated in designing a blockchain/NFT-based employee honor system and implemented data backup and recovery mechanisms to support system stability and data integrity.

Built a deep learning algorithm for automatic modulation recognition by combining invariant scattering features, wavelet denoising, residual modules, SSN, and LSTM to improve recognition stability under low-SNR conditions.

• Compared hard/soft thresholding and Sqtwolog, Rigrsure, Minimaxi, and Heursure wavelet-denoising strategies, adopting Rigrsure hard-threshold denoising to improve robustness under low-SNR conditions.
• On the Radio 2018.01A public dataset, the algorithm exceeded 80% accuracy at 6dB and reached 90.12% at 10dB and 91.26% at 20dB.

Produced digital-asset research on NFT financialization paths and NFTFi liquidity mechanisms, later published in the CDEMS 2024 proceedings.

• Analyzed how Blend improves NFT liquidity through peer-to-peer lending, perpetual loan terms, and automatic refinancing.
• The research was recommended for CDEMS 2024, published by Atlantis Press, and indexed by CPCI, CNKI, and Google Scholar.