Yiran HuMathematician & Developer

Ambees

• We help small companies hire college students through a 10-minute AI conversation, not a resume.
• Built the product end to end with Next.js, React, TypeScript, Neon PostgreSQL, and Vercel.
• Built the AI voice interviewer on OpenAI Realtime that runs the 10-minute onboarding conversation and feeds the matching pipeline.
• Rebuilt candidate vetting so an LLM turns a recruiter's ask into structured criteria, a deterministic ranker scores every candidate, and a second model checks the top 12. Every ranking change must pass a test set of synthetic candidates, including decoys and keyword-stuffers, before it ships.
• Completed the Georgetown Summer Launch Incubator and pivoted to a recruiter marketplace after customer discovery.

UC Berkeley Sky Computing Lab

• Research on multi-agent LLM systems with the FrontierCS group, studying whether letting teams of LLM coding agents communicate helps them beat a single agent at the same token budget.
• Built the FrontierCS 2.0 task pipeline for our multi-agent harness (Codex, DeepSeek).

Data Discovery Program, UC Berkeley Data Science

• Designing LLM-as-a-judge scoring for Berkeley course tutors (Data 8, Data 100, E127), focusing on self-judge bias.
• Measuring judge-human agreement with Cohen's kappa on OpenRouter and NRP models (Qwen3, gpt-oss).
• Writing Python automated checks that flag answer leakage and missed student mistakes in tutor responses.
• Read 10+ LLM tutor benchmarking papers (MRBench, MathTutorBench) and presented findings to the team.

UC Berkeley School of Education

• Co-authoring two papers in preparation for Sociology Compass and Organization Theory.
• Coauthored a 27-page paper on how the 12 largest U.S. school districts govern generative AI, published in Educational Policy (second author).
• Synthesized 30+ empirical studies on organizational structure and built a cross-sector case base for four human-AI work arrangements (tool, assistant, co-worker, supervisor).
• Received a donor-funded $3,500 summer research stipend.

Computer Science Mentors (CSM)

• Lead weekly review sections for CS 88 students on Python, recursion, and data abstraction.
• Record video walkthroughs of past exam problems used by current and future cohorts.

UC Berkeley Department of Mathematics

• Investigated Nash equilibria in Brownian Boost tug-of-war games with resource allocation dynamics.
• Built a web version of the finite game (Trail of Lost Pennies) and compared human play traces to ABMN equilibria.

Data Discovery Program, UC Berkeley Data Science

• Benchmarked small models (Qwen2.5, TinyLlama) on accuracy, bias, and compute with HuggingFace Transformers and llama.cpp.
• Built Jupyter evaluation pipelines and designed a RAG curriculum module for a future Berkeley data science course.

Stanford Mathematics Camp

• Formalized the Rubik's Cube CFOP method as a subgroup chain decomposition and authored a 20-page paper that received a top evaluation.