Brian Zhou Liu

Product Development Intern | Remote

• Engineered an OCR bulk upload feature using Gemini Flash Lite, allowing suppliers to upload thousands of unstructured product sheets (PDFs, Excel, etc.) to be sorted into the Asakana MongoDB database.
• Conceptualized and developed a DialogflowCX agent to automatically process buyer orders (identifying buyers, applying specific pricing policies, propagating orders) using Cloud SQL and Twilio SMS.
• Conducted market research with industry insiders to translate insights into actionable product improvements.

Research Intern under Prof. Aobo Li | La Jolla, CA

• Benchmarked various ML architectures (vision transformers, 2D convolutions, graph-based DeepSphere, 3D autoencoders) against KamNet for neutrinoless double beta decay detection, evaluating background rejection accuracy and tuning hyperparameters.
• Applied RESuM surrogate modeling framework to COH-Ar-750 detector simulation data. Used multi-fidelity approach (conditional neural process + Gaussian process) to assess neutron background rates, demonstrating optimal shielding configurations that reduce need for expensive Monte Carlo simulations.
• "Efficient optimization of COHERENT detector design parameters with the Rare Event Surrogate Model (RESUM)," submitted to Machine Learning and the Physical Sciences, NeurIPS 2025.

Research Intern under Prof. Ran Xiao | Atlanta, GA

• Evaluated 10 data compression methods using XResNet for heart attack diagnostics. Identified 12-lead format for optimal performance (0.926 AUC) and autoencoder for efficiency.
• Constructed anatomically informed feature tensor for XResNet, achieving significant (p<0.001) AUC increase to 93.7% for myocardial infarction detection. Developed regional learners (inferior, septal, anterior, lateral) combined with global learner using 1D CNNs to capture cross-lead spatial patterns.
• Zègre-Hemsey, J. K., Ding, C., Liu, B., Wright, D. W., et al. "Enhancing Deep Learning in Detecting Acute Myocardial Infarction via Anatomically Informed 12-Lead ECG."

Sushi Hackathon at Stanford University | Stanford, CA

• Engineered a dual-purpose carpal tunnel syndrome (CTS) wearable diagnostic brace using an Arduino Nano, dual load cells, and FSRs.
• Developed a Python-based Bluetooth API to transmit sensor data wirelessly for real-time analysis.
• Contributed to a full-stack dashboard with RAG AI agent for insights and data visualization of 50,000+ data points.

Berkeley AI Hackathon | San Francisco, CA

• Integrated Google Custom Search JSON API and Gemini Flash 2.0 to search public records from sign-up data.
• Engineered user risk prediction system with Claude Sonnet 4 to generate transparent 0–100 risk scores.
• Created React/Supabase dashboard with email notifications for high-risk users.

ACM x Intel Hackathon | Santa Clara, CA

• Developed an interactive knowledge graph organizing academic works and relationships.
• Implemented full RAG system allowing users to query uploaded scientific literature with citations.
• Optimized UI with D3 graph visualizations and integrated note-taking.

UCSD DiamondHacks | La Jolla, CA

• Engineered prediction module fusing Gemini-generated descriptions with FDA database API.
• Fine-tuned Gemini Flash 2.0 with Vertex AI for regulated pill descriptions.
• Deployed full-stack app on Vercel and Google Cloud Run with custom domain.

High School Research | Great Neck, NY

• Developed assistive computer vision system using YOLOv4-tiny, achieving 87.34% mAP.
• Engineered Raspberry Pi 4B headwear, reducing costs by 90% vs commercial alternatives.
• Implemented real-time audio feedback system using Python/OpenCV.