DISCOVER PROJECT
Immersive Video Coaching for CPAP (Continuous Positive Airway Pressure) Adherence
Edge-to-Cloud AI Pipeline Delivering Personalized Medical Coaching Videos Under 60 Seconds via 5G/6G Quality-on-Demand
Designed and built a 6G Quality on Demand (QoD) enabled video generation pipeline that uses Raspberry Pi edge computing to detect real-time CPAP telemetry anomalies and automatically deliver personalized, medically accurate coaching videos to patients within strict network latency constraints
PYTHON - FASTAPI - TENSORFLOW LITE - RASPBERRY PI - GOOGLE VERTEX API - CAMARA QOD API - MATPLOTLIB - WEBVTT - SQLITE - UVICORN - 5G/6G EDGE COMPUTING
Project Overview
Built a full data-to-impact pipeline connecting CPAP sensor telemetry (airflow, SpO2, mask pressure) to an edge-deployed anomaly classifier on a Raspberry Pi, which triggers a cloud orchestration layer to select, adapt, or generate personalized coaching videos. To prevent AI "hallucination" of medical data, the architecture decouples clinical fact from generative animation, Python renders deterministic charts from real sensor values first, and Google Vertex API only animates that verified chart, while separate WebVTT files guarantee 100% accurate bilingual (English/French) subtitles. Delivery is coordinated through a CAMARA 5G Quality-on-Demand session targeting sub 60 second latency for cached assets and under 15 minutes for freshly generated clips
The system follows four stages: an edge node (Raspberry Pi) running a quantized TFLite classifier on sensor windows, a cloud orchestration server that maps confirmed deviations to one of three coaching paths (retrieve, adapt, or generate a video), a generative video pipeline combining Matplotlib charts with Veo animation, and a bilingual HTML5 dashboard that logs patient feedback for model retraining. This "Generative Fallback" architecture was formalized so that even unmapped or rare anomalies receive a fast, deterministic, and medically safe video response within the SLA window
