Show HN: 通过鼠标移动检测访客情绪(情绪推断)
1 分•作者: sentientiq•8 个月前
我构建了一个系统,通过鼠标遥测技术实时检测访客情绪——无需调查,无需追踪像素,零个人身份信息(PII)。
问题:
你的分析告诉你发生了什么(用户跳出),但不知道为什么(他们感到困惑、沮丧或因为价格太高而放弃)。
工作原理:
- JavaScript 捕获鼠标移动、点击模式、滚动行为
- 情绪推断引擎(Claude Sonnet)分析行为特征
- 系统检测:沮丧、困惑、犹豫、自信、退出意图
- 语境感知干预措施在毫秒内部署
- 反馈循环从结果中学习
技术栈:
- EC2 上运行的 20 个微服务(情绪推断、跨行业机器学习、干预引擎)
- NATS 用于实时消息流
- Supabase 用于持久化存储
- 经过速率限制和生产环境加固
与众不同之处:
- 无需调查(实时行为推断)
- 无 PII(仅情绪状态,不跟踪身份)
- 空间感知(干预措施与页面上下文匹配)
- 自我改进(从转化结果中学习)
演示:
访问 <a href="https://sentientiq.ai" rel="nofollow">https://sentientiq.ai</a> - 你会感受到它的作用。交互式演示展示了我们检测到的内容。
技术深度解析:
在 <a href="https://sentientiq.ai" rel="nofollow">https://sentientiq.ai</a> 上打开浏览器控制台并观察:
遥测流(鼠标移动、点击、模式)
情绪检测(好奇 → overwhelmed → 自信)
干预措施部署(上下文响应)
完整架构:20 个微服务,NATS 流,Claude 推断(Haiku→Sonnet 升级),速率限制为 10 Sonnet 调用/分钟/会话。详细文档即将发布。欢迎在此处解答技术问题。
我独自一人花了 6 个月时间构建了这个系统。差点死了两次。很乐意收到来自 HN 社区的反馈。
查看原文
I built a system that detects visitor emotions in real-time from mouse telemetry - no surveys, no tracking pixels, zero PII.<p>The Problem:
Your analytics tell you what happened (user bounced), but not why (they were confused, frustrated, or priced out).<p>How it works:
- JavaScript captures mouse movements, click patterns, scroll behavior
- Emotional inference engine (Claude Sonnet) analyzes behavioral signatures
- System detects: frustration, confusion, hesitation, confidence, exit intent
- Context-aware interventions deploy in milliseconds
- Feedback loop learns from outcomes<p>The Stack:
- 20 microservices on EC2 (emotional inference, cross-vertical ML, intervention engine)
- NATS for real-time message streaming
- Supabase for persistence
- Rate-limited and hardened for production<p>What makes this different:
- No surveys (real-time behavioral inference)
- No PII (emotional states only, no identity tracking)
- Spatial awareness (interventions match page context)
- Self-improving (learns from conversion outcomes)<p>Demo:
Visit <a href="https://sentientiq.ai" rel="nofollow">https://sentientiq.ai</a> - you'll feel it working on you. The interactive demo shows what we detect.<p>Technical Deep Dive:
Open the browser console on <a href="https://sentientiq.ai" rel="nofollow">https://sentientiq.ai</a> and watch:<p>Telemetry stream (mouse movements, clicks, patterns)
Emotion detection (curiosity → overwhelm → confidence)
Intervention deployment (contextual responses)
Full architecture: 20 microservices, NATS streaming, Claude inference (Haiku→Sonnet escalation), rate-limited to 10 Sonnet calls/min/session. Detailed docs coming soon. Happy to answer technical questions here.<p>Built this solo over 6 months. Nearly died twice. Would love feedback from the HN crowd.