I build reproducible AI tools for multimodal evaluation and browser-agent regression. My projects connect runnable code, independent outcome checks, and inspectable results.
Projects and results · Engineering case studies · About AlvenX
| Project | What you can inspect | Read the engineering case |
|---|---|---|
| BrowserAgentRegression | A local browser-agent lab with independent DOM checkpoints and controlled interface changes. Its frozen deterministic reference driver passed 36/36 task/variant attempts. | Browser lifecycle and independent scoring |
| OpenMultimodalLab | Two pinned VLM backends evaluated on 102 synthetic tasks, with three measured repetitions per backend and task on a recorded RTX 4060 Laptop. | Multimodal measurement on an 8GB GPU |
| ReproLock | An experimental Playwright verifier binding reviewed tests, exact application revisions, build inventories, and cleanup. The DrawDB case includes an ordinary Playwright control. | Verified before and after browser regressions |
The articles explain the problem, implementation choices, verification and measured limits. Source repositories retain the original evidence and reproduction instructions. Projects are developed with AI tooling; the published records distinguish generated candidates, controlled experiments and independently checked outcomes.
- PhysGauge: controlled collision worlds for checking how video metrics respond to known physical violations.
- GeometryAudit: recorded foundation-geometry runs and independent relative-pose calculations. The original reliability-adapter candidate was retired after a prior-art review; its evidence remains available.
- SpatialSceneLab: a first scene-package experiment with editable authored objects, explicit units and frames, GLB round-trip checks and a separate retained-reconstruction coordinate audit. English and Chinese stage reports describe its measured limits.
ReproLock's test repair experiment separates candidate test outcomes from independent business-state observations, with authored controls and one recorded model proposal replay.
Research status, input conditions and current limits are shown alongside each result on alvenx.com.
我围绕多模态评测与浏览器 Agent 回归开发可复现的 AI 工具，将可运行代码、独立验收和结果记录连起来。
建议先读三个工程案例：
这些文章说明遇到的问题、实现选择、验证和测量范围；源码仓库保留原始证据与复现步骤。开发中使用 AI 工具，生成的候选、控制实验与独立核验结果分别记录。
SpatialSceneLab 中文阶段报告记录对象编辑、单位与坐标导出控制；ReproLock 中文研究记录说明测试修复后业务错误检测能力的实验口径。