01
2026 · rPPG 评测 · 诊断 AgentrPPG evaluation · diagnostic agent
Geneva / rppgbench.onlineGeneva / rppgbench.online
我独立搭建了从模型提交、Docker 沙箱推理到指标和排行榜的 rPPG 隐藏评测平台。测试集包含 120 段 MMPD 视频,按光照、运动和肤色组织结果,并与 5 类 rPPG 基线比较。双语诊断 Agent 可读取提交、指标和 Docker 日志,解释失败或低分原因。I built the rPPG hidden-evaluation platform end to end, from model submission and Docker-sandboxed inference to metrics and leaderboards. Its test set contains 120 MMPD clips, with results organized by lighting, motion, and skin tone and compared with five rPPG baselines. A bilingual diagnostic agent reads submission data, metrics, and Docker logs to explain failed or low-scoring runs.
- 评测链路覆盖分片上传、解包、构建、推理、打分和入库;平台展示 MAE、RMSE、Pearson、CCC 及分场景误差。The evaluation path covers chunked upload, unpacking, build, inference, scoring, and persistence, with MAE, RMSE, Pearson, CCC, and condition-specific errors.
- 基于 DeepSeek 的双语诊断 Agent 可只读调用提交状态、评测指标、Docker 日志和 FAQ,用多步工具调用解释模型为什么失败或分数偏低。A bilingual DeepSeek diagnostic agent can call read-only tools for submission status, evaluation metrics, Docker logs, and FAQ retrieval to explain failed runs or unexpectedly low scores.
- Agent 工具限定在当前用户的作用域内,并设置步数上限与超时;查不到数据时不生成虚构结论。Agent tools are user-scoped and bounded by step and time limits; empty tool results do not become fabricated diagnoses.
02
2026 · 生成式交互 · 家庭研究Generative interaction · field study
StoryEcho 原型StoryEcho prototype
我负责把家长记录的进食行为接入孩子后续故事,并加入阅读与互动日志、失败重试和端到端检查。该原型支持了 11 个家庭的五日探索性研究。I worked on the path from a parent’s meal record to the child’s next story, then added reading and interaction logs, retry handling, and end-to-end checks. The prototype supported a five-day exploratory study with 11 families.
03
2024 · 生理信号 · Qt · 端侧处理Physiological signals · Qt · edge processing
光纤非接触式夜间高血压测量原型Fiber-optic prototype for nocturnal hypertension sensing
这是一个枕下光纤生理感知原型,目标是为夜间高血压筛查提供连续的心率与呼吸率观测。我负责信号处理、端侧推理优化和 Qt 系统实现。This under-pillow fiber-optic prototype explores continuous heart-rate and respiratory-rate sensing for nocturnal hypertension screening. I worked on signal processing, edge-inference optimization, and the Qt application.
- 系统以约 100 Hz 通过串口 / 蓝牙采集,支持实时绘图、在线 / 离线回放、导出和异常标记;处理包括 IIR / FIR 及 FFT / STFT。The system acquires serial or Bluetooth data at about 100 Hz and supports real-time plots, online/offline replay, export, and anomaly marking, with IIR/FIR and FFT/STFT processing.
- 实验室静息条件下心率 MAE 为 3–5 BPM。这是原型阶段的实验结果,不表示已完成临床验证。Heart-rate MAE was 3–5 BPM under laboratory rest conditions. This is a prototype-stage result, not a claim of clinical validation.