IPBMS — Real-time Patient Monitoring

The problem

Falls and seizures need a fast response, and caregivers cannot watch every camera all the time. Automatic detection only helps if staff trust it: a system that raises false alarms all day gets muted.

What I built

IPBMS is a real-time patient monitoring system for Whammy Tech. From June 2025 to January 2026 I was AI Team Lead, leading four engineers across backend, frontend and mobile, and I owned all the AI work:

• Fall and seizure detection on live RTSP camera streams using YOLO pose estimation.
• LLM-written incident reports, so a caregiver gets a readable account of what happened, not just a raw alert.
• A five-frame rule — an alert fires only after the event is detected in five frames in a row.
• The detection feeds a real-time event dashboard, live alerts in the mobile app, a web admin portal for patients, cameras and caregivers, and an AI summary of each patient's recent events.

How it works

Camera streams are read over RTSP and processed frame by frame. YOLO pose estimation finds people and their body keypoints, and the pose over time is used to spot a fall or a seizure. A single odd frame is not enough to alert: the event has to hold for five consecutive frames, which filters out the brief misdetections that cause most false alarms. When an alert does fire, an LLM writes the incident report that goes to caregivers.

Outcome

The five-frame rule cut false alarms by 84% across 4,000 test cases. The system ran on-site on two real cameras.

Stack

• YOLO pose estimation, OpenCV
• RTSP camera streams
• LLM-written incident reports
• Python