Qianyu Zheng

Cumulative GPA: 4.00/4.00

Cumulative GPA: 4.00/4.00

Built agentic systems for automated model development in the physical sciences.
- Shipped an agentic hyperparameter optimization framework for machine learning interatomic potentials and property-prediction models into internal research infrastructure; tuned models set internal state-of-the-art benchmarks and reached production use.
- Developed a multi-agent auto-research system that evolves MLIP architectures by reasoning over model design, proposing new architectures concurrently with in-flight optimization runs for higher experimental throughput than sequential loops.
- Validated the system by recovering strong-baseline performance from an underperforming architecture and by extending a model to a property outside its original scope.

Prepared multi-instrument field measurements for wake model validation in the AWAKEN project.
- Processed a year of TB-scale, multi-instrument wind measurement data (dual-Doppler radar, lidar, disdrometer, met stations) with NetCDF and Spark pipelines to curate benchmark cases.
- Invented a local contrast filter to discard cases with unphysical spikes in radar data, identified meteorological and seasonal correlates of the artifacts, and validated that flagged cases showed degraded wind speed and angle accuracy against reference instruments.
- Benchmarked two engineering wake models across the full AWAKEN domain (two wind farms, > 100 km²) under varying turbulence regimes. First-author paper in JRSE.

Applied machine learning and graph analysis to large protein families.
- Explored protein families as sequence similarity graphs with embedding-distance-weighted edges, evaluating graph sparsification, community detection, and centrality metrics to identify functionally important sequences in the cytochrome c and influenza families.
- Designed a cluster-based, distance-maximizing data split strategy for non-i.i.d. protein sequences, combining ESM embedding clustering with a surrogate-model-assisted evolutionary search; outperformed random splits and SpanSeq on downstream classification.

Advisor: Victor Fung

Researching machine learning for materials discovery and simulation in the Fung Group.
- Developed a physics-informed pretraining strategy for machine learning interatomic potentials, validated across three chemistries (silica, phosphorus, MPTrj) and three architectures (CGCNN, TorchMD, M3GNet); reduced unphysical molecular dynamics artifacts and improved recovery of AIMD trajectories over regularization baselines. First-author paper in JCIM.
- Built a differentiable geometry toolbox exposing structural properties as autograd-optimizable losses and served it to LLM agents over MCP, reaching 87.5% success on 40 natural-language structure-editing tasks; designed a propose-act-feedback multi-agent system that tripled target band gaps on 9/10 inverse-design tasks.
- Leading an agentic descriptor-mining project that applies SISSO within an agent framework to discover interpretable descriptors for superconducting critical temperature across ~8,000 conventional superconductors.

Led a 2-person team building a multimodal tool for free-text queries over protein sequences in the UniProt database.
- Trained a CLIP-style dual encoder (ESM2 + BERT) on a 10k-pair corpus curated by pairing UniProt entries with LLM-generated layman and professional descriptions.
- Reached 94.05% top-50 accuracy on layman queries (vs. 67.33% for an NER/keyword baseline) and cut query latency from over 24 hours to under 40 seconds.
- Deployed a quantized inference service with Flask and Docker on AWS (ECS/Fargate, ECR, Route 53), live at nl2prot.org.

Participated in the Workout Of the Day (WOD) prediction project group at Data Science @ GT.
- Used Python to build data cleaning and feature engineering pipelines for the downstream machine learning tasks.
- Leveraged modern optimization libraries to design an automated hyperparameter search pipeline for modeling.

Implement a pipeline to extract source data from various forms of scientific plots.
- Utilized YOLO and fineuned OCR to extract data from various forms of scientific plots.
- Deploy with Flask backend and HTML/CSS frontend.

A tool tailored for PR teams of companies to analyze impact of tweets on stock prices.
- Trained a word2vec model for tweet analysis.
- Utilized GPT for generation of sample tweets.
- Deploy with Flask backend and Wix on Google Cloud.

Instructor for the 2024 iteration of a free summer bootcamp introducing ~200 science-major students to scientific computing and deep learning with Python and PyTorch. Led the lecture on deep learning and PyTorch fundamentals (autodiff, building MLPs from scratch), and assisted on the lecture covering graph neural networks for molecular applications.

Supported a large-enrollment (~500 students per semester) introductory course on object-oriented programming and JavaFX, working with a team of ~25 TAs across four semesters on grading, office hours, and weekly recitations. Served as Q&A forum lead for two semesters, owning homework and exam logistics announcements and coordinating TA responses.

Funded assistantship supporting full-time research as a Master's student in the Fung Group, Georgia Tech

Competitive, faculty-mentored research award supporting undergraduate research at Georgia Tech

Awarded each semester for achieving a 4.0 GPA; received every semester of Bachelor studies

Expertise in Manage workbook options and settings, Manage and format data, Create advanced formulas and macros, and Manage advanced charts and tables in Excel.

Expertise in Manage worksheets and workbooks, Manage data cells and ranges, Manage tables and table data, Perform operations by using formulas and functions, and Manage charts in Excel.

Advanced certification validating the ability to design, build, and deploy ML solutions on AWS.

Foundational certification validating knowledge of AWS Cloud concepts, services, security, and pricing