Barry Si-Qi Wu | 伍思琦

Email   |   GitHub   |   LinkedIn   |   Scholar

Profile

Hi there, I'm Siqi, currently an algorithm engineer at Shopee. I obtained my bachelor's degree in AI from the School of Computer Science and Engineering at UESTC. My current research interests involve Online Advertising, Quantitative Investing, AI-Medicine and ML System～

Connect with me via email at dionysusfhs@gmail.com or via WeChat. All advice and questions are welcome!

Education

University of Electronic Science and Technology of China B.Eng. in AI, School of Computer Science and Engineering 2021.09 - 2025.06

The First High School of Changsha, Hunan 2018.09 - 2021.06

• PGMV Model: Consolidated separate regional models into a unified model, reducing maintenance overhead. Introduced query-item interaction features, optimized the model architecture, and removed dependency on organic pCR (org_pcr, an upstream prediction score), enabling a full rollout across all regions. Refined the definitions of price and sold_cnt features to improve prediction accuracy.


• Uplift Model: Developed a multi-treatment coupon uplift model based on DragonNet. Applied FiLM to incorporate merchant/platform voucher context features and item-level price-effect sensitivity features. Further upgraded voucher cost estimation to a dispatch-to-redeem modeling framework, improving the offline voucher_cost PCOC from approximately 0.6 to 0.9.

• Explored mixed-precision algorithms to solve high-precision linear least squares on Tensor Cores; analyzed singular-value distributions to address convergence on ill-conditioned matrices.
• Implemented double-blocking TSQR with JIT kernels and achieved up to 2.2× speedup over cuSOLVER for FP64 QR decomposition; overall solver up to 26× speedup on favorable matrices.
• Trained a 2-layer LSTM on 15,000 matrix instances to classify matrix types and select robust solver strategies (99.6% classification accuracy).

• Installed and indexed MIMIC-III/IV datasets and developed retrieval pipelines using LLMs for RAG-enhanced clinical prediction.
• Designed a BERT-based classification algorithm and LLM summarization strategies to reduce token overflow and improve performance.

• Built cross-modal retrieval using BLIP and CLIP, implemented a web interface and demonstrated robust retrieval performance.

• Developed an automated essay scoring pipeline using DeBERTa; applied AdamW, CosineAnnealingWarmRestarts, and K-fold CV; achieved MCRMSE=0.4492 on Kaggle.

[1] Gaoyuan Z., Siqi Wu, and Shaoshuai Z.. “Improving the Reliability on Scientific Computing Using Tensor Cores: A Case Study on Solving High Precision Linear Least Square Problems.” (Submitted to SC2025)

[2] Hansheng Wang et al. “Improving Tridiagonalization Performance on GPU Architectures.” (Accepted by PPoPP2025)

• 2026 Tencent Ads Algo Competition: Top 3%
• 2025 Baidu Business AI Technology Innovation Competition: Top 20
• 2023 Huawei Intelligence Foundation Scholarship

Languages & Tools: C, C++, Python, MATLAB; Linux, PyTorch, TensorFlow, Pandas, NumPy, LaTeX, HTML, SQL

Languages: Mandarin (Native), English (IELTS 7.0)

Interests: Bartending, Badminton, Long-distance running, Reading, Music, Movies, Photography

Template adapted from original.
Links: GitHub • Email