What it is
Bicycle Radar turns an iPhone mounted on your bike into a rear-view safety radar. On-device computer vision watches the road behind you, and when a car or truck approaches you get an audio alert and a vibration, so you never need to look at a screen while riding. It records rides with detection overlays if you want the receipts, estimates distance and speed for what it sees, and runs entirely on the phone: no account, no cloud, no uploaded video.
I built the original prototype after too many close passes on my own weekend rides. In 2026 it grew into a shipping product: the 1.0.25 release, with detection rebuilt and verified on-device, went to App Review in July.
How it works
- A TensorFlow Lite detector (EfficientDet-Lite0) runs against the live camera feed on the phone; alerts key off distance, speed, and approach vector
- Getting iOS detection solid was a proper debugging saga: model loading inside a background isolate, bi-planar camera formats, and a broken quantized model export each masqueraded as the same symptom, and each got its own fix and regression guard
- A bundled demo clip plays through the real detector, boxes and alerts included, so a curious rider (or an App Store reviewer) can watch it work without mounting a phone on a bike
- Privacy is structural: camera frames are analyzed and discarded in real time, location feeds only the on-screen speed, and the privacy policy is a static page that says exactly that
Why it matters
Safety systems only work if people believe them. That means low false alarms, but it also means honesty about what the system is doing with your camera and your location. Tuning for trust, not just raw accuracy, is the thread that carried from the first prototype into the shipping app. Built end to end with Claude Code.