Personal project · 2024 · shipped

Sleep Vision

A bedside camera that learns how you actually sleep.

Apple Watch biometrics and a live video feed, fused on a sub-100ms timeline, with an LLM that compares each night to published sleep research and tells you what changed.

Apple Watch heart data synced to video within 100ms
PyTorch + FFmpeg pipeline for real-time feature extraction
LLM-generated nightly feedback grounded in research

PyTorch
FFmpeg
Swift

<100msvideo ⇄ watch sync

100%on-device

how it works

01
Capture
An overhead camera and Apple Watch stream in parallel; heart-rate and motion are fused with the video on a sub-100ms shared timeline.
02
Extract
A PyTorch + FFmpeg pipeline pulls real-time features from the feed — movement, position, restlessness — frame by frame.
03
Explain
An LLM compares each night against published sleep research and writes plain-language feedback on what actually changed.

up next · chapter 06Cyber SimLabLearn cybersecurity by living the incident — hands-on attack & defense simulations, not slideshows.

Sleep Vision

Capture

Extract

Explain