Summary：How I Trained an AI to Generate My Cycling Highlights Automatically **Introduction** Every weekend

referrerpolicy="no-referrer"
style="max-width:100%;height:auto;display:block;margin:0 auto;">

How I Trained an AI to Generate My Cycling Highlights Automatically

**Introduction**
Every weekend I strap a GoPro to my helmet and log ride data with a Garmin .fit file. After a long climb the footage piles up—hours of raw video that sit untouched on my hard drive. Turning those clips into a punchy 60‑second highlight reel used to mean endless scrubbing, manual cuts, and painstaking telemetry overlays. I decided to automate the process, training an AI model to scan, rank, and stitch the best moments while burning in speed, heart‑rate, and elevation data. The result is a repeatable workflow that costs roughly $0.04 per ride and delivers a polished video in minutes.

**Key Developments**
The pipeline begins with two inputs: the GoPro MP4 files and the corresponding .fit telemetry export. Using Google’s Gemini‑3.5‑Flash model, I built a custom prompt that instructs the AI to analyze each second of video for visual interest—sharp turns, summit views, group surges—and to cross‑reference that with the .fit metrics for spikes in power or heart‑rate. The model outputs a ranked list of clip candidates, each with a confidence score and suggested