How I Built It

Everything flows through a single RunSummary model. Whether the source is the baked-in demo or an uploaded activity, every run is normalized into the same structure—distance, pace, heart rate, cadence, per-kilometer splits, best efforts, PR count, and detected race events—before anything touches an API.

Parsing real running files

Supporting both .gpx and .tcx turned out to be the trickiest engineering challenge.

TCX files provide cumulative distance at every point, which is essential because treadmill workouts often contain no GPS coordinates at all. GPX files, on the other hand, only contain latitude and longitude, so distance has to be reconstructed using the haversine formula between consecutive points.

Multi-lap TCX files introduce another wrinkle by resetting their cumulative distance every lap, so those segments have to be stitched back together manually. Regardless of format, the raw activity never leaves the browser.

Turning data into commentary

Gemini receives the summarized run data—splits, heart-rate progression, detected surges, PRs, and pacing trends—and writes a structured 45–60 second sports broadcast.

The prompt guides the narration through a consistent flow:

• hype introduction


• the grind


• the redline moment


• the closing kick


• the payoff

If a run has no heart-rate data or no PRs, Gemini naturally adapts instead of mentioning missing information.

Giving it a voice

I originally chose a voice from ElevenLabs' shared Voice Library, only to discover that library voices require a paid subscription for API access—even on free accounts.

The fix was switching to one of the built-in premade voices ("Charlie — Deep, Confident, Energetic"), which works perfectly on the free tier. It was a small lesson, but exactly the kind of issue that can quietly break a live demo.

Protecting the free tier

ElevenLabs' free tier has a very limited generation quota, so I designed the app around it.

• The homepage demo plays a pre-generated MP3 committed as a static asset, meaning unlimited playback without API usage.


• Live generations are protected by Upstash Redis with both a per-IP hourly limit and a deployment-wide lifetime budget.


• If a generation fails midway, its reserved quota is refunded so transient failures don't permanently consume the available budget.

Stack: Next.js 16, React 19, TypeScript, Tailwind CSS v4, Framer Motion, @google/genai, ElevenLabs REST API, Upstash Redis.

Running is already emotional.

I wanted the recap to sound the way it felt.

Prize Categories

Submitting for:

• 🧠 Best Use of Google AI — Gemini transforms real running telemetry into an original sports commentary instead of a templated summary.


• 🔊 Best Use of ElevenLabs — The voice isn't just an output; it's the experience that turns data into something worth listening to.