Zhexiang (Bruce) Li

M.S. in Computer Science

B.S. in Computer Science, Magna Cum Laude · GPA 3.76 / 4.00

Advisor: Prof. Yuzhe Yang

• Working on multimodal foundation models for diabetes that forecast glucose from continuous glucose monitoring together with insulin, meals, activity, and patient context.
• Harmonized CGM data from 49 datasets into a common format with meal, insulin, and wearable context.
• Building patient-disjoint MetaboNet benchmarks to compare GluFormer, T1DFormer, and custom predictive-state and Transformer forecasters at 30-, 60-, and 120-minute horizons.

Research with a Ph.D. student in Prof. Yan Liu's group · Project poster

• Studied SARS-CoV-2 antiviral discovery through reproducible benchmarks of molecular descriptor-based models, Chemprop, and hybrid approaches across seven Therapeutics Data Commons datasets, using scaffold splits.
• Analyzed descriptor importance and SHAP attributions to examine the molecular features behind model predictions.

Zhexiang Li (sole author). Under review

Designed lesion-supervised token routing after a frozen vision backbone so that dermoscopy classifiers keep diagnostic regions under a fixed token budget. Trained it with budget curriculum learning, diversity regularization, and attention distillation, and evaluated it on ISIC 2019.

Under review at AAAI 2027

Represented mass-conserving fragment pairs as sparse evidence tokens for spectrum retrieval. Used closest-mass and overlap-hard negatives, alongside evidence-removal tests, to examine whether the exported evidence affects ranking.

Zhexiang Li, Haoyu Wang, Yutong Bao, David Woodruff. Under review · arXiv:2505.11040

First-author work advised by Prof. David P. Woodruff (CMU). Proposed pre-scoring and implemented informative-key selection in Triton; integrated it with HyperAttention and evaluated it on ChatGLM with 131k-token contexts.

S. Li, Z. Li, Z. Liu. Advances in Economics, Management and Political Sciences, 16(1), 357–366. DOI

Advised by Prof. Ming Gu (UC Berkeley). Led a team modeling game-release decisions as two-player normal-form games and enumerating Nash equilibria with Python and Gambit.

• Designed short-video recommendation using bidirectional user-video interest propagation, exposure budgets across multiple timescales, and fatigue controls to manage repeated recommendations.
• Developed 11-language audio-visual affect inference with evidence fusion and a fixed-point decision layer that can abstain or report conflicting modalities; implemented in C for Cortex-M3 and ARM926EJ-S targets.
• Screened sub-10 mm piezoelectric robot architectures and specified a PZT-steel cantilever experiment for a 9 mm candidate; work concluded at experiment and procurement planning.

• Built Shopify-integrated pages and Google Maps features; improved responsive layouts, accessibility, SEO, and maintainable content structures.

• Led product and engineering for AI-assisted housing debt collection, including data pipelines and automated workflows.
• Set technical direction, hiring priorities, and delivery standards during startup development.

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

Gathered requirements for a NetEase CodeWave internal platform, helped recruit engineers, and contributed to program structure and interface planning.

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

Supported MedTech deal screening, market analysis, due diligence, financial modeling, and valuation.

Next.js · TypeScript · Supabase · LLM and vision models · howtoeat.food

• Built and launched an AI kitchen web app that plans daily and weekly meals and grocery lists from each user's inventory, leftovers, preferences, cookware, and cooking time.
• Added photo-based inventory recognition and LLM recipe import with user corrections, admin AI review, and deduplication; isolated each user's data with Supabase Postgres row-level security.
• Built an offline multimodal pipeline (subtitles, ASR/OCR, keyframe vision) that turns cooking videos into structured recipes for the library.

Python · FastAPI · MySQL · RAG · WeChat Mini Program

• Built and deployed the FastAPI backend on WeChat Cloud Hosting for a clinic's Mini Program and web workspace covering clinical teaching, clinician collaboration, and patient intake.
• Designed five-role access control so LLM-organized visit records stay drafts until a doctor confirms them; added an asyncio queue that schedules LLM calls under global, per-user, and per-API limits.
• Implemented RAG with BGE embeddings, ChromaDB cosine search, and a keyword fallback.

Python · Claude Code

Built an agent skill that researches a company from its name or pitch deck and drafts an evidence-backed investment report; a Python QA harness checks each section's sources, depth, and completeness.

Swift · SwiftUI · Vision parsing · OCR · Source code

Built photo-to-contact extraction with OCR fallback, multi-photo review, bilingual records, and searchable people and company profiles.

Prof. Lise Getoor

• Held office hours and discussion sessions on logistic regression, SVMs, and tree models.
• Co-designed Jupyter Notebook assignments and contributed to the coding autograder.

VEX Robotics & Olympiad in Informatics

Supported robot design and algorithmic problem-solving instruction, scrimmages, and events.