Suzy (Ziwei) Su苏紫薇 Suzy Su

My work我的工作

I developed a consultation-practice prototype that separates the rubric judge from the simulated patient. Code controls fact access; every awarded mark links back to the dialogue that earned it.

我开发模拟问诊练习原型，将评分裁判与模拟病人分离。代码控制事实访问，每个得分链接到对应对话证据。

Compared quality, latency and cost to select the judge. On one synthetic case, 24 adversarial prompts produced 1.33 leaked facts per attack in the full-context control and none in the guarded user-visible output. Per-session and daily spend caps limit model spend.

比较质量、延迟与成本后选择裁判。单个合成病例的 24 次攻击中，全量上下文对照组每次泄露 1.33 个锁定事实，守卫后的用户可见输出未观测到泄露。每局与每日设置模型花费上限。

Exact-line accuracy: the predicted set of rubric items matches the labelled set for a line. Scope: one synthetic case; no clinician review. These results do not establish clinical validity or general safety.精确行准确率：每句话预测的评分项集合与标注集合完全一致的比例。范围：单个合成病例，未经临床审核。上述结果不代表临床有效性或普遍安全性。

My work我的工作

I built a public job-search workflow with AI coding assistants. An LLM reviewer scores a CV against a five-dimension rubric with a truthfulness veto, in a separate session from the model that wrote it; rule-based checks flag unsupported numbers, skill terms and metric abbreviations against a fact base; they do not verify every claim in context. Command-line tools check the UK sponsor register. 190 tests at 97% coverage, CI on Ubuntu and Windows with Python 3.10 and 3.12.

我借助 AI 编码助手开发公开求职工作流。LLM 评审按五个维度的评分标准给简历打分，「真实性」一项可一票否决，而且在和写稿模型分开的会话里运行；之后用规则检查事实库中找不到的数字、技能词和指标缩写；它不能逐句验证完整语义。另有查询英国担保名册的命令行工具。190 个测试、覆盖率 97%，CI 跑在 Ubuntu 和 Windows、Python 3.10 和 3.12 上。

My work我的工作

I built black-box UI testing tools for Android builds that UIAutomator and Appium can't reach. pytest and ruff run in GitHub Actions across Python 3.9–3.12, and the README states the framework's limits, including 50–500 ms OCR latency.

我为 UIAutomator 和 Appium 读不到的 Android 版本开发黑盒界面测试工具。GitHub Actions 跑 pytest 和 ruff，覆盖 Python 3.9–3.12；README 写明了框架的局限，包括 50–500 ms 的 OCR 延迟。

My contribution · team project我的贡献 · 团队项目

A two-person project with Xiaomin Fan, awarded “Best Product Integration with Z.ai” at VibeHack London 2026. I contributed to game design, frontend development, testing and integration. The project uses Next.js, TypeScript and Zod, with 128 tests. We used Claude Code on GLM; the game runs on pre-written branches and deterministic logic.

与 Xiaomin Fan 合作的双人项目，获 VibeHack London 2026「Best Product Integration with Z.ai」。我参与玩法设计、前端开发、测试与集成；项目使用 Next.js、TypeScript 和 Zod，共有 128 个测试。借助 Claude Code（GLM 模型）开发，游戏运行预写分支与确定性逻辑。