UC Berkeley Bachelor of Computer Science, 2021-2025 · Google AI Intern, 2025 · University of Melbourne MCS, 2026-present
一心不乱 / One mind, undisturbed.
Focused execution, rigorous evaluation, and trustworthy AI systems.
AI研究と剣道に通じる集中と修練。
以剑道的专注，打磨 AI 系统与评测。
I am an AI researcher interested in how language models behave when they become systems: agents that use tools, retrieve context, evaluate themselves, and operate inside real workflows. My work focuses on practical evaluation, agent reliability, multilingual safety, and transformer-based NLP systems.
- Education: UC Berkeley, Computer Science undergraduate, 2021-2025; University of Melbourne, Master of Computer Science, 2026-present
- Experience: Google AI, AI Intern, 2025
- Research taste: careful benchmarks, readable systems, reproducible experiments
- Practice: Japanese kenshi, guided by 一心不乱: disciplined attention under pressure
| Area | What I build |
|---|---|
| LLM evaluation | Reproducible eval pipelines, rubric graders, model comparison reports |
| AI agents | Trace analysis, tool-use evaluation, reliability and failure-mode studies |
| AI safety | Multilingual safety tests, refusal/over-refusal analysis, risk taxonomies |
| Transformers | Small model experiments, attention visualization, NLP training notes |
| Project | Signal |
|---|---|
| hms-harmful-brain-activity-classification | Kaggle HMS EEG classification Silver Medal solution archive, 123rd of 2,767 teams |
| agent-safety-eval-lab | Flagship lab for agent traces, tool-call grading, and safety evaluation |
| mcp-tool-security-playground | Tool-use security, permission policies, and prompt-injection threat modeling |
| rag-eval-observatory | RAG observability with retrieval, faithfulness, and failure-case analysis |
| multilingual-llm-safety-bench | English/Japanese/Chinese safety mini-benchmark and model behavior report |
| ISC-Bench-Reproduction | Reproduction-oriented work around LLM safety and agent evaluation |
| Transformers-Projects | Transformer/NLP project lab and experiments |
| Competition | Result |
|---|---|
| HMS - Harmful Brain Activity Classification | Kaggle Silver Medal, 123rd / 2,767 teams |
| Repository | Purpose |
|---|---|
| transformer-from-scratch-notes | Attention, tiny tokenizer, and mini training-loop notes |
| llm-eval-cookbook | Exact match, rubric, preference, pairwise, and JSON-schema eval recipes |
| agent-trace-viewer | Lightweight HTML trace viewer for agent messages, tools, and failures |
| prompt-robustness-suite | Prompt versioning, A/B testing, and failure clustering |
| open-model-benchmark-cards | Structured benchmark-card generator for open models |
English · 日本語 · 中文
- Email: yutoterashima4@gmail.com