About

Current

I am an ML Research Scientist at Sapient Intelligence in Beijing. I am a co-first author of Praxist, an autonomous AI-scientist framework with DAG-based execution lineage for reproducible solution provenance (arXiv, GitHub). On the 75-task MLE-bench, Praxist won 60 medals (49 gold) against 55 (34 gold) for Claude Code with Opus, at about a twelfth of the cost. If you find Praxist useful, please give it a star on GitHub!

In 2026, I showcased Praxist at the AI4 conference through live demonstrations and conversations with prospective customers and collaborators.

I am leading a joint project between Sapient and Shanghai Jiao Tong University (with Prof. Liang Hong) that applies Praxist to the Virtual Cell Challenge 2026: zero-shot prediction of single-cell responses to CRISPRi gene knockdowns. Our team is currently ranked 187th of 1,320 (top 14.2%) on the preliminary leaderboard.

Since April 2026, I have been collaborating with Prof. Jian Tang at Mila on Bayesian optimization for auto-research systems, where a Gaussian process predicts which candidate solutions are worth expanding and Thompson sampling chooses the next experiment.

Research

Based at the University of Cambridge and collaborating with Shanghai Jiao Tong University, I developed PLASMA, a plug-and-play optimal-transport module that adds local substructure alignment to any protein language model. In detecting shared substructures (motifs, binding sites, and active sites), PLASMA reached ROC-AUC 0.96–0.99 versus 0.81–0.93 for the widely used global aligner TM-align at about 50× its speed, and outperformed the state-of-the-art local method EBA at about 3× its speed, with much cleaner alignments. It was accepted as a poster at ICLR 2026.

At Cambridge, I built Topotein, a complete SE(3)-equivariant framework that adds a secondary-structure level to protein representations via a hierarchical hypergraph. This coarser-grained view gives a better global understanding of protein structure: it achieved the best SCOP fold classification accuracy (43.3%, vs. 38.4% for the same network without the hierarchy). I presented this work in an invited talk at Cambridge, and the project received the Department of Computer Science and Technology’s Highly Commended M.Phil Project Prize 2024–2025.

I also built MoRE-GNN, a heterogeneous graph autoencoder that learns relational graphs directly from single-cell RNA, protein, and ATAC data; it outperforms MOJITOO on strongly correlated CITE-seq data (BM-CITE ARI 0.892 vs. 0.868) and enables cross-modal prediction. With Prof. Hatice Gunes, I developed GraphAU-Pain, presented at the IJCAI 2025 MiGA Workshop, and extended it into GraphAU-Pain++, a data-efficient AU-graph transfer learning framework published in IEEE Consumer Electronics Magazine.

At UCL, I collaborated with clinicians from UCL Hospital on readmission prediction using remote patient monitoring data. I developed a SQL-based data pipeline, performed clinical feature engineering, and built readmission-prediction models with a ROC-AUC of 0.79 ± 0.03, using SHAP to show which clinical features drive each prediction.

Education

At the University of Cambridge, I completed an M.Phil. in Advanced Computer Science with Distinction, ranked 8th out of 60. Courses included Geometric Deep Learning, Affective AI, Mobile & Wearable ML, NLP, and Multi-agent RL.

I graduated with First-Class Honors in B.Sc. Computer Science from University College London. Courses included Machine Learning, Computer Vision, Reinforcement Learning, Mathematics & Statistics, and Software Engineering.

Industry Experience

At iFLYTEK Healthcare, I built a knowledge-distillation pipeline that uses a teacher LLM to synthesize medical instruction data for fine-tuning the model behind iFLYTEK’s Xiaoyi AI health-assistant app, raising its instruction-following accuracy from 34% to 82%.

At Luojin Data Information, I developed a RAG pipeline with a MongoDB vector store and LangChain agents for automated financial-report classification, achieving 92% query-translation accuracy and a 97% entity-matching rate.

Through UCL’s industrial collaboration program with the NHS, I built a system that crawls a hospital’s website and auto-generates a homepage navigation assistant that answers visitor questions and guides them to the right pages, cutting setup from weeks of manual development to 30 minutes. The project was demonstrated live at Great Ormond Street Hospital and presented to IBM, Microsoft, and Intel.

Academic Service

I am a reviewer for NeurIPS 2026 and for the AI4Science Workshops at NeurIPS 2025 and ICML 2026. I was also a Programming Tutor at UCL in 2022–23.

Research Interests

I am interested in AI agents, large language models, AI for biology and health, and geometric deep learning. I welcome opportunities to collaborate on technically ambitious, high-impact work.

Technical Skills

Machine Learning: PyTorch, PyTorch Lightning, PyTorch Geometric, NumPy, Pandas, Scikit-learn

Languages: Python, JavaScript, Java, C++, R

Research Infrastructure: SLURM, Weights & Biases, TensorBoard