Data Engineer — VLM Evaluation
Video Rebirth · Hong Kong, China
- Owned the end-to-end architecture and delivery of the company's video evaluation platform (Next.js + PostgreSQL + Prisma), spanning annotator assignment, Arena-style model comparisons, self-service package creation, and a permission-aware RBAC model
- Established a disciplined local → staging → production release path and migrated staging to Alibaba Cloud ACK; sustained production usage across 35+ users, 24+ evaluation packages, and 21,000+ evaluation items with zero P0 incidents
- Engineered an automated collection pipeline across 8+ video-model families, unifying prompt dispatch, asynchronous generation monitoring, validation, and OSS ingestion to produce 5,600+ evaluation videos
- Re-architected evaluation asset governance through a zero-loss OSS migration, credential-vault hardening with AES-256-GCM controls and reveal-rate limiting, and standardized archival for new datasets
- Developed reusable evaluation methodology assets, including a VQA-derived eight-subdimension rubric, a 120-prompt first/last-frame suite across three new dimensions, and a competitive intelligence base covering seven World Model companies
Data Engineer – 视频模型评测
Video Rebirth · 中国 – 香港
- 主导公司视频评测平台从架构设计到生产交付（Next.js + PostgreSQL + Prisma），覆盖标注任务分配、Arena 双模型对战、XLSX prompt suite 自助建包及权限解耦的 RBAC 体系
- 建立 local → stg 回归 → prod 的标准化发布链路，并推动 stg 迁移至阿里云 ACK；平台稳定支撑 35+ 用户、24+ 评测包与 21,000+ 评测条目，保持零 P0 事故
- 搭建覆盖 8+ 视频模型家族的自动化采集 pipeline，统一 prompt 分发、异步生成监控、结果校验与 OSS 入库，累计产出 5,600+ 评测视频
- 重构评测资产治理体系，零丢失完成 OSS 结构化迁移，并通过 AES-256-GCM 审计、凭证泄露限流及数据集自动归档强化安全与可追溯性
- 构建可复用的评测方法论资产，包括基于 VQA benchmark 的八子维度评分体系、覆盖三个新增维度的 120 条首尾帧 prompt suite，以及七家 World Model 公司的竞品知识库
