Yuma Watanabe

I’m an AI coding-agent evaluator and red-teamer focused on adversarially testing AI coding agents and large language models. I design complex, multi-step and multi-turn instructions that probe reasoning, code changes, and tool use—especially where failures only surface under real difficulty and complexity.

In my contract role (remote), I directed AI coding agents to edit real, git-cloned repositories to expose weaknesses. I prioritize cross-file and cross-component consistency, identifying cases where edits look correct in isolation but break connections between components, then reproduce and document failure modes clearly so engineering teams can act on them.

I also build my own machine-learning models in Python, and I’ve applied that mindset to quantitative finance. In an independent project, I built a risk-focused swing-trading model using the moomoo/futu API, emphasizing minimizing maximum drawdown and validating with a 10-year backtest including look-ahead/data-leakage checks.

I’m a native Japanese speaker with B2-level English, and I enjoy multilingual adversarial evaluation where communication quality matters. I’m drawn to teams that value rigorous test design, structured technical documentation, and safety improvements grounded in measurable failure-mode evidence.