Oct 2025 – Present· Researcher (co-first author)
Carnegie Mellon University · Prof. Min Xu & Prof. Rebecca E. Taylor
An LLM agent framework that takes a scientist's intent all the way to executed experiments on liquid-handling robots, with checks before a run and automatic recovery when something goes wrong.
- Designed a multi-agent framework covering experiment planning, protocol verification and real-time monitoring across multiple liquid handlers.
- Built pre-execution validation and a hybrid rule + LLM tool-calling recovery agent that repairs failures during a run.
- Created a 48-task real-experiment planning suite and a 96-scenario fault-injection simulator; benchmarked against 8 LLMs and agent baselines.
Jul 2026 – Sep 2026· Remote Research Assistant (co-author)
Northeastern University · Dr. Tianyu Liu & Prof. Wengong Jin
A benchmark that tests whether AI agents can operate real scientific software on a virtual desktop. I built its chemistry domain, the largest and hardest part of the benchmark.
- Built the chemistry task set, 43 tasks (about 30% of the benchmark), each with an automatic RDKit grader for SMILES, InChIKey and stereochemistry.
- Connected the tasks to the OSWorld agent harness and evaluated frontier models.
- Chemistry turned out to be the hardest domain; the best model solved only about 44% of its tasks.
Oct 2025 – Jul 2026· Research Assistant (first author)
Carnegie Mellon University · Prof. Min Xu
Segmenting proteins in real cryo-electron tomograms without any real voxel labels, by adapting models trained on simulated data while avoiding the collapse that sparse foregrounds usually cause.
- Trained 3D protein segmentation models on simulated cryo-ET and adapted them to real tomograms with no real labels.
- Developed a collapse-free adaptation method (foreground-balance loss, Fourier style alignment, prototype pseudo-labels) that beats the best baseline by 1.8× and reaches 78% of the fully supervised upper bound.
- Built a benchmark of 9 baselines across backbones and datasets; our method was also the most stable across random seeds.
Feb 2023 – May 2023· Undergraduate Researcher (co-first author)
Xi'an Jiaotong-Liverpool University · Prof. Zhen Wei
A multi-task CNN that denoises MeRIP-seq signals from low sequencing depth and calls m6A modification peaks.
- Preprocessed and analyzed HEK293 MeRIP-seq data with ExomePeak2.
- Co-developed the CNN-based multi-task model for signal denoising and m6A peak calling.
- Contributed to manuscript writing and data visualization (under review at BMC Bioinformatics).
Jan 2023 – Oct 2024· Team Leader & Software Developer
XJTLU-Software iGEM team · advised by Guoxia Han & Dechang Xu
A retrieval-augmented question-answering system for synthetic biology built on a fine-tuned LLM. Gold Medal in the iGEM AI & Software track.
- Designed a synthetic biology RAG search framework on top of a fine-tuned LLM.
- Fine-tuned Llama 3.1 8B on domain data, with embedding-based retrieval, tokenizer optimization and evaluation modules.
- Built a benchmark for question answering in synthetic biology.
- iGEM 2024 Gold Medal, AI & Software track.
Jan 2022 – Oct 2022· Website Group Leader
XJTLU-CHINA iGEM team · advised by Guoxia Han & Dechang Xu
An engineered biofilm reactor for domestic sewage purification. Gold Medal and nominated for Best Environment Project at iGEM 2022.
- Designed the biofilm reactor instrument for the water purification system.
- Built biophysical models with GROMACS molecular dynamics, Michaelis–Menten kinetics and the law of mass action.
- Led the team website and illustrated the full project.
- iGEM 2022 Gold Medal, nominated for Best Environment Project.