Solution (What we built)

• Built a real-time video pipeline that ingests IP camera streams and runs on-device/server-side inference continuously.
• Developed a pose-based behavior classifier to differentiate “potentially threatening” vs “non-threatening” motion patterns (e.g., aggressive movements, sudden escalation cues).
• Implemented alerting + event capture: short clips + timestamps surfaced for quick review.

Outcome

• Won Fan Favorite at Virginia Tech Kickstart.
• Demoed end-to-end system: stream ingestion → inference → alert/clip review (link below).

Tech

(OpenCV / PyTorch / YOLO or pose model / RTSP / Docker / AWS

https://youtu.be/IZjgUgAPve4

KickStart VT Visualization.pdf

RefiBot Brochure (1).pdf

Features Page.docx.pdf