CV

Lead Developer — NuFold Multimer: Generative Modeling for RNA Complexes

Kihara Lab, Purdue University

• Extended a Transformer-based RNA structure model from single-chain to multi-chain prediction, integrating a diffusion module to generate multiple candidate 3D structures.
• Built a data-curation and teacher-inference pipeline, running inference on approximately 20,000 RNA complexes; integrated multiple data sources, corrected inconsistent labels and identifiers, and applied confidence-based filtering for training.
• Ran matched training ablations using 8 NVIDIA A100 80GB GPUs across 4 nodes to evaluate diffusion-decoder stability.
• Enabled training crops of approximately 1,000 tokens by integrating NVIDIA cuEquivariance and a memory-efficient diffusion implementation.
• Benchmarked AlphaFold3 and Boltz-2 on RNA complexes to compare prediction accuracy and characterize model limitations.

Evaluation & Benchmarking — CryoASTRA: Cryo-EM Atomic Structure Modeling

Kihara Lab, Purdue University

• Built a Python/SLURM benchmarking pipeline for CryoASTRA and seven baselines across 103 test cases, with resumable execution, validated result caching, and input/code provenance tracking.
• Standardized prediction-to-reference matching and coverage-aware metrics for incomplete outputs; implemented matched-target comparisons and bootstrap confidence intervals to quantify performance differences.
• Evaluated five generated samples per input; benchmark results showed a mean TM-score of 0.972 and 92.5% atom-placement accuracy within 3 Å across 101 protein-containing targets, with a statistically significant TM-score gain of 0.117 over CryoAtom2.

Lead Developer — Distance-AF Multimer: Constraint-Guided Protein Complex Modeling

Kihara Lab, Purdue University

• Implemented per-target test-time optimization to adapt pretrained model predictions to user-supplied geometric constraints without retraining the underlying model.
• Designed a coarse-to-fine optimization strategy to improve convergence from inaccurate initial predictions and reduce structural artifacts.
• Built a PyTorch optimization workflow supporting structures with 50,000+ atoms, parallelizing independent target optimizations across GPUs.
• Evaluated on 27 challenging multimer targets, 8 peptide complexes, and 6 assemblies with 5+ chains; achieved lower RMSD than Chai-1, Protenix, and AlphaLink2 on 22/26, 19/27, and 26/27 benchmark targets, respectively, with up to 57.6 Å lower RMSD than AlphaFold-Multimer.
• Conducted ablations across 11 complexes with 0–30 distance constraints and uncertainty of ±2.5 Å and ±5 Å; retained acceptable docking accuracy on 8/11 targets under both uncertainty settings.
• First-author manuscript in preparation.

Designer & Sole Developer — DAQ-Refine: Cryo-EM Protein Model Refinement Service

Kihara Lab, Purdue University

• Independently built and deployed a public React/Flask application with interactive Mol* visualization; integrated into EMSuite and used for approximately 10–20 refinement jobs per month.
• Migrated a Colab workflow to lab infrastructure and automated SLURM job orchestration to run three refinement strategies per protein chain, select the best results, and assemble the final structure.
• Developed reusable input-validation, early error-reporting, and progress-tracking components for integration with other EM-server algorithms.
• Diagnosed recurring job failures reported by lab researchers and external users; added optional sequence alignment, specific input-error messages, and input-preparation documentation to help users resolve incompatible inputs.

Research Assistant

MSSN Lab, Purdue University

Advisor: Prof. Chunyi Peng

• Independently developed the initial mperf platform using Django and Android for remote experiment control and automated cellular-network measurements; collected 705 hours of data comprising 4.6 million throughput samples across three U.S. carriers, supporting two published studies.
• Redesigned MMMap’s database and visualization workflows for 32,000+ network cells and 18,700+ handoff events, achieving a 10× speedup in data analysis and display.
• Collected and analyzed 45 hours of UAV network measurements over 1,010 km, using radio-access-network traces to compare airborne and ground-level throughput, latency, and mobility failures.
• Applied lightweight formal analysis to 3GPP Release 17 IoT protocols, identifying four vulnerabilities and developing 12 proof-of-concept demonstrations.

Teaching Assistant — CS536: Computer Networking

Purdue University

Selected semesters

• Led office hours and debugging sessions, authored assignments and exam questions, and graded across multiple cohorts — explaining complex networking systems and protocols to students.

PhD

Purdue University

Computer Science

• Expected May 2027
• Advisor: Prof. Daisuke Kihara
• Master of Science in Computer Science, May 2025

BS

Huazhong University of Science and Technology (HUST)

Computer Science and Technology

• GPA: 3.8/4.0
• National Scholarship (2/120)
• Freshmen Scholarship (1/120)

National Scholarship

HUST

Ranked 2/120

Freshmen Scholarship

HUST

Ranked 1/120

Model Student of Academic Records

HUST

Ranked 5/120

Outstanding Student

HUST

Ranked 4/120