CV
A PDF version of the CV can be downloaded by clicking the icon on the right.
Contact Information
| Name | Yan Zhou (Terry) |
| terryzhou [at] fas [dot] harvard [dot] edu | |
| Location | Cambridge, MA 02138 |
| GitHub | https://github.com/tz1211 |
| https://www.linkedin.com/in/yan-zhou-a9085b256/ |
Education
-
Sep 2025 - May 2027 Cambridge, US
Master of Science
Harvard University
Data Science
-
Sep 2022 - Jun 2025 London, UK
Bachelor of Science
London School of Economics and Political Science
Politics and Data Science
Experience
-
Nov 2025 - Present Cambridge, US
AI Researcher
ML Foundations Group, Harvard SEAS
Investigating generalization behavior of boomerang-distillation on post-trained LLMs.
- Built model capability evaluation pipelines and optimized model distillation training workflow for multi-GPU cluster.
- Demonstrated that post-trained teacher capabilities can be zero-shot transferred via layer patching onto base student models.
-
Sep 2025 - Present Cambridge, US
AI Researcher
AI4LIFE Lab, Harvard SEAS
Research on activation steering for reasoning control in LLMs.
- Demonstrated that steering for specific model personas can lead to more concise model CoT and enhanced performance on reasoning benchmarks.
-
Sep 2025 - Nov 2025 Cambridge, US
Technical AI Safety Fellow
Harvard AI Safety Student Team, Harvard SEAS
Participated in advanced AI safety reading groups.
-
Jun 2024 - Jun 2025 London, UK
Research Assistant
Data Science Institute, LSE
Led development of a RAG chatbot for queries regarding LSE policy and regulations.
- Designed automated data ingestion and embedding pipelines.
- Built scalable backend for institutional policy QA.
-
Jun 2023 - Oct 2024 London, UK
Research Assistant
Data Science Institute, LSE
Built ML framework to identify factors that drive MP rebellions in UK Parliament.
- Utilized speech embeddings to estimate political ideological differences, achieving a 20% increase in F1-score for predicting voting behavior.
-
Apr 2023 - Apr 2024 London, UK
Data Science Intern
The Tecsa Group
Projects
-
Are Personas All You Need? Stress-Testing Persona Vectors as Alignment Tools
Evaluation of persona vectors as alignment tools in LLMs.
- Evaluated the reliability of persona vectors as low-dimensional alignment tools for modeling behavioral traits.
- Analyzed geometric properties of persona vectors to quantify cross-trait interference during finetuning.
- Identified key robustness failures such that minor natural language variations produce persona vectors with cosine similarity as low as 0.6.
-
DataDetox
An interactive AI agent orchestration system for ML model and data provenance.
- Developed an agentic AI system to trace model and training data lineage, allowing ML practitioners to identify hidden upstream risks and enhance model and data provenance.
- Engineered an asynchronous multi-agent backend that retrieves disparate metadata and queries from graph database to construct model and data lineage trees.
- Deployed a production-ready stack using Docker and Kubernetes for scalability.
Skills
Programming Languages: Python (Advanced), R (Advanced), Typescript (Beginner), SQL
Machine Learning: PyTorch, Transformers, Huggingface, Scikit-learn
ML Engineering / Infrastructure: vLLM, Finetuning, Model Distillation, Quantization, Agents, OpenAI API, RAG, Vector Databases, MongoDB, Neo4j, Docker, Kubernetes, Weights & Biases, Git, CI/CD, FastAPI, GCP, AWS, Slurm
Data & Visualization: NumPy, Pandas, SciPy, Matplotlib, Plotly, Seaborn, Jupyter
Languages
English : Native/Bilingual
Chinese : Native/Bilingual
Spanish : Intermediate