experience

• Developed ConvoLearn, a 40K-turn post-training dataset for strengthening dialogic capabilities of LLMs in tutoring contexts.
• Directed large-scale human data operations (750+ teachers across provisioning, annotation, and evaluation): recruitment, task design, qualification gating, and pilot protocols for 20-turn dialogues aligned to a pedagogical framework.
• Designed a multi-stage QA pipeline with hybrid human-LLM annotation, inter-rater reliability analysis, safety filtering, and expert adjudication, producing a high-quality, SFT-ready dataset.
• Conducted an ecological validity analysis linking the dataset to meaningful signal on real-world educational outcomes.
• Fine-tuned multiple 7–8B open-weight LLMs (Mistral-7B-Instruct, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct) using QLoRA with progressive multi-turn sampling.
• Built intrinsic and human evaluation pipelines (RoBERTa-based effectiveness model and blinded teacher study), benchmarking against a proprietary state-of-the-art model.
• Led the research program across platform development, data operations, modeling, evaluation, grants, paper writing, and public dataset release on Hugging Face.

Stanford HAI ConvoLearn (arXiv)

• Joint industry-lab work on AI-based simulation of personas and their evaluation, grounding persona models in real datasets and designing human-in-the-loop mechanisms for co-design and controllability.

Stanford HAI IBM Research

• Developed a RAG system for prerequisite assessment generation across three courses in Stanford’s introductory Biology sequence.
• Integrated SME review into item generation and piloted assessments over three quarters (N > 500 students).
• Performed IRT-based refinement of prerequisite assessments, including validation, reliability, and multiple forms of validity analysis, to inform the department about factors influencing student performance.

Stanford Biology

• Developed a large-scale binary annotated dataset (12,000+ instances) of feedback from West Africa on actionability (actionable vs. vague).
• Developed a semi-supervised technique to build the dataset: rubric grounded in academic literature on feedback, manual coding of a subset, inter-annotator agreement analysis, and RoBERTa fine-tuning on the hand-labeled data to scale labeling.
• Performed interpretability analysis to quantify trends in linguistic features of actionable feedback from the classified dataset; paper published at the 20th ACL BEA workshop.

BEA 2025 paper

• Specialization in natural language processing and measurement.
• Awarded the Stanford HAI RA position: 50% tuition + stipend.
• Awarded travel scholarships to ACL 2025 (Vienna) and ASU+GSV 2026 (San Diego).
• Awarded the Stanford-Peking Fellowship on Generative AI + Healthcare for a two-week bootcamp in Beijing and Shenzhen.
• Graduate Specialist for Outcomes at Stanford's Office for Inclusion, Belonging & Intergroup Communication.

Stanford GSE