Zhexiang (Bruce) Li

• Designed a short-video recommender with bidirectional user-video interest propagation and separate session, bridge, and long-term exposure budgets. Added fatigue controls that cap repeated topics while reserving traffic for new interests.
• Built 11-language audio-visual affect inference with synchronization checks, presentation-attack detection, evidence fusion, and a fixed-point Q15 decision layer that can abstain or flag modality conflicts.
• Validated the recommendation logic with 629 tests and the multimodal pipeline with 139 Python tests, type and contract checks, and cross-compilation for Cortex-M3 and ARM926EJ-S.
• Evaluated a 9 mm piezoelectric millirobot concept and translated mechanical constraints into a PZT-steel cantilever test plan with explicit pass/fail criteria.

• Built the data and workflow architecture for an AI-assisted housing debt-collection platform; set technical priorities and delivery standards.

• Built Shopify-integrated pages and Google Maps features; fixed responsive, ADA compliance, and SEO issues.

• Built an LLM-assisted pipeline to ingest, clean, and cluster 5,000+ medical-malpractice documents for quality-of-care review.
• Wrote experiment checklists and acceptance tests for vascular-care laboratory workflows.

• Validated product firmware and built an ESP32-to-Alibaba Cloud pipeline that published I2C sensor data as MQTT topics for internal dashboards.

• Screen MedTech opportunities and support market research, due diligence, financial modeling, and valuation.

Advised by Prof. David P. Woodruff, Carnegie Mellon University | arXiv:2505.11040

• Proposed pre-scoring to select informative key intervals before attention; implemented Triton kernels and benchmarked against HyperAttention and LevAttention.

Research with Prof. Yan Liu's group

• Benchmarked RDKit, Chemprop, and hybrid models on seven TDC datasets with scaffold splits; ran feature-importance and SHAP analyses. Extra Trees reached AUROC 90.0 ± 6.4 on SARSCoV2_3CLPro_Diamond.